Yihang Yao, Guangtao Zeng, Raina Wu, Yang Zhang, Ding Zhao, Zhang-Wei Hong, Chuang Gan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The emergence of Large Language Models (LLMs) with strong reasoning capabilities marks a significant milestone, unlocking new frontiers in complex problem-solving. However, training these reasoning models, typically using Reinforcement Learning (RL), encounters critical efficiency bottlenecks: response generation during RL training exhibits a persistent long-tail distribution, where a few very long responses dominate execution time, wasting resources and inflating costs. To address this, we propose TLT, a system that accelerates reasoning RL training losslessly by integrating adaptive speculative decoding. Applying speculative decoding in RL is challenging due to the dynamic workloads, evolving target model, and draft model training overhead. TLT overcomes these obstacles with two synergistic components: (1) Adaptive Drafter, a lightweight draft model trained continuously on idle GPUs during long-tail generation to maintain alignment with the target model at no extra cost; and (2) Adaptive Rollout Engine, which maintains a memory-efficient pool of pre-captured CUDAGraphs and adaptively select suitable SD strategies for each input batch. Evaluations demonstrate that TLT achieves over 1.7x end-to-end RL training speedup over state-of-the-art systems, preserves the model accuracy, and yields a high-quality draft model as a free byproduct suitable for efficient deployment. Code is released at https://github.com/mit-han-lab/fastrl.

Multi-task learning (MTL) presents greater optimization challenges than single-task learning (STL) due to conflicting gradients across tasks. While parameter sharing promotes cooperation among related tasks, many tasks require specialized representations. To balance cooperation and specialization, we propose Mod-Squad (Chen et al. 2023), a modular transformer-based model composed of a "squad" of experts. Each task activates a sparse subset of experts through a differentiable matching process, guided by a novel mutual information-based loss. This modular structure avoids full backbone sharing and scales effectively with the number of tasks and dataset size. In this extended version, we generalize Mod-Squad to support multi-dataset pre-training, enabling joint learning across disjoint, single-task datasets (e.g., ImageNet, COCO, ADE20 K). This is achieved via a new formulation of the mutual information loss that unifies learning across heterogeneous sources. More importantly, while most prior work in large models has focused on efficiency, few have explored adjustable efficiency. In this study, we further evaluate the model's generalization to downstream tasks and introduce a set of efficient adaptation techniques that leverage Mod-Squad's modularity for flexible fine-tuning-enabling dynamic adjustment of model size, parameter count, and computational cost. Additionally, we present a hybrid adaptation scheme that combines these techniques to achieve favorable performance-efficiency trade-offs. In summary, Mod-Squad provides a robust foundation for sparse modular models that can learn from diverse supervision and datasets. Its emergent modularity enables strong generalization, decomposition into high-performing components, and rapid, resource-efficient adaptation for downstream applications.

We introduce UniMuMo, a unified multimodal model capable of taking arbitrary text, music, and motion data as input conditions to generate outputs across all three modalities. To address the lack of time-synchronized data, we align unpaired music and motion data based on rhythmic patterns to leverage existing large-scale music-only and motion-only datasets. By converting music, motion, and text into token-based representation, our model bridges these modalities through a unified encoder-decoder transformer architecture. To support multiple generation tasks within a single framework, we introduce several architectural improvements. We propose encoding motion with a music codebook, mapping motion into the same feature space as music. We introduce a music-motion parallel generation scheme that unifies all music and motion generation tasks into a single transformer decoder architecture with a single training task of music-motion joint generation. Moreover, the model is designed by fine-tuning existing pre-trained single-modality models, significantly reducing computational demands. Extensive experiments demonstrate that UniMuMo achieves competitive results on all unidirectional generation benchmarks across music, motion, and text modalities.

Constructing compact and informative 3D scene representations is essential for effective embodied exploration and reasoning, especially in complex environments over extended periods. Existing representations, such as object-centric 3D scene graphs, oversimplify spatial relationships by modeling scenes as isolated objects with restrictive textual relationships, making it difficult to address queries requiring nuanced spatial understanding. Moreover, these representations lack natural mechanisms for active exploration and memory management, hindering their application to lifelong autonomy. In this work, we propose 3DMem, a novel 3D scene memory framework for embodied agents. 3D-Mem employs informative multi-view images, termed Memory Snapshots, to capture rich visual information of explored regions. It further integrates frontier-based exploration by introducing Frontier Snapshots—glimpses of unexplored areas—enabling agents to make decisions by considering both known and potential new information. To support lifelong memory in active exploration settings, we present an incremental construction pipeline for 3D-Mem, as well as a memory retrieval technique for memory management. Experimental results on three benchmarks demonstrate that 3D-Mem significantly enhances agents’ exploration and reasoning capabilities in 3D environments, highlighting its potential for advancing applications in embodied AI.

Research on 3D Vision-Language Models (3D-VLMs) is gaining increasing attention, which is crucial for developing embodied AI within 3D scenes, such as visual navigation and embodied question answering. Due to the high density of visual features, especially in large 3D scenes, accurately locating task-relevant visual information is challenging. Existing works attempt to segment all objects and consider their features as scene representations. However, these task-agnostic object features include much redundant information and missing details for the task-relevant area. To tackle these problems, we propose LSceneLLM, an adaptive framework that automatically identifies task-relevant areas by leveraging LLM’s visual preference for different tasks, followed by a plug-and-play scene magnifier module to capture fine-grained details in focused areas. Specifically, a dense token selector examines the attention map of LLM to identify visual preferences for the instruction input. It then magnifies fine-grained details of the focusing area. An adaptive self-attention module is leveraged to fuse the coarse-grained and selected fine-grained visual information. To comprehensively evaluate the large scene understanding ability of 3D-VLMs, we further introduce a cross-room understanding benchmark, XR-Scene, which contains a series of large scene understanding tasks including XR-QA, XR-EmbodiedPlanning, and XR-SceneCaption. Experiments show that our method surpasses existing methods on both large scene understanding and existing scene understanding benchmarks. Plunging our scene magnifier module into the existing 3D-VLMs also brings significant improvement. Code and data are available at https://github.com/Hoyyyaard/LSceneLLM

Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as VCA. Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency.

Diffusion models have shown promise in data-driven planning. While these planners are commonly employed in applications where decisions are critical, they still lack established safety guarantees. In this paper, we address this limitation by introducing SafeDiffuser, a method to equip diffusion models with safety guarantees via control barrier functions. The key idea of our approach is to embed finite-time diffusion invariance, i.e., a form of specification consisting of safety constraints, into the denoising diffusion procedure. This way we enable data generation under safety constraints. We show that SafeDiffusers maintain the generative performance of diffusion models while also providing robustness in safe data generation. We evaluate our method on a series of tasks, including maze path generation, legged robot locomotion, and 3D space manipulation, and demonstrate the advantages of robustness over vanilla diffusion models.

Large language models (LLMs) have demonstrated remarkable capabilities across a range of text-generation tasks. However, LLMs still struggle with problems requiring multi-step decision-making and environmental feedback, such as online shopping, scientific reasoning, and mathematical problem-solving. Unlike pure text data, collecting large-scale decision-making data is challenging. Moreover, many powerful LLMs are only accessible through APIs, which hinders their fine-tuning for agent tasks due to cost and complexity. To address LLM agents' limitations, we propose a framework that can automatically learn a reward model from the environment without human annotations. This model can be used to evaluate the action trajectories of LLM agents and provide heuristics for task planning. Specifically, our approach involves employing one LLM-based agent to navigate an environment randomly, generating diverse action trajectories. Subsequently, a separate LLM is leveraged to assign a task intent and synthesize a negative response alongside the correct response for each trajectory. These triplets (task intent, positive response, and negative response) are then utilized as training data to optimize a reward model capable of scoring action trajectories. This reward model can be integrated with LLM-based agents and various planning algorithms to enhance task-solving performance. The effectiveness and generalizability of our framework are demonstrated through evaluations conducted on different agent benchmarks. In conclusion, our proposed framework represents a significant advancement in enhancing LLM agents' decision-making capabilities. By automating the learning of reward models, we overcome the challenges of data scarcity and API limitations, potentially revolutionizing the application of LLMs in complex and interactive environments. This research paves the way for more sophisticated AI agents capable of tackling a wide range of real-world problems requiring multi-step decision-making.

Tracking dense 3D motion from monocular videos remains challenging, particularly when aiming for pixel-level precision over long sequences. We introduce DELTA, a novel method that efficiently tracks every pixel in 3D space, enabling accurate motion estimation across entire videos. Our approach leverages a joint global-local attention mechanism for reduced-resolution tracking, followed by a transformer-based upsampler to achieve high-resolution predictions. Unlike existing methods, which are limited by computational inefficiency or sparse tracking, DELTA delivers dense 3D tracking at scale, running over 8x faster than previous methods while achieving state-of-the-art accuracy. Furthermore, we explore the impact of depth representation on tracking performance and identify log-depth as the optimal choice. Extensive experiments demonstrate the superiority of DELTA on multiple benchmarks, achieving new state-of-the-art results in both 2D and 3D dense tracking tasks. Our method provides a robust solution for applications requiring fine-grained, long-term motion tracking in 3D space.

We present TopoGaussian, a holistic, particle-based pipeline for inferring the interior structure of an opaque object from easily accessible photos and videos as input. Traditional mesh-based approaches require tedious and error-prone mesh filling and fixing process, while typically output rough boundary surface. Our pipeline combines Gaussian Splatting with a novel, versatile particle-based differentiable simulator that simultaneously accommodates constitutive model, actuator, and collision, without interference with mesh. Based on the gradients from this simulator, we provide flexible choice of topology representation for optimization, including particle, neural implicit surface, and quadratic surface. The resultant pipeline takes easily accessible photos and videos as input and outputs the topology that matches the physical characteristics of the input. We demonstrate the efficacy of our pipeline on a synthetic dataset and four real-world tasks with 3D-printed prototypes. Compared with existing mesh-based method, our pipeline is 5.26x faster on average with improved shape quality. These results highlight the potential of our pipeline in 3D vision, soft robotics, and manufacturing applications.

In this paper, we investigate the problem of embodied multi-agent cooperation, where decentralized agents must cooperate given only egocentric views of the world. To effectively plan in this setting, in contrast to learning world dynamics in a single-agent scenario, we must simulate world dynamics conditioned on an arbitrary number of agents' actions given only partial egocentric visual observations of the world. To address this issue of partial observability, we first train generative models to estimate the overall world state given partial egocentric observations. To enable accurate simulation of multiple sets of actions on this world state, we then propose to learn a compositional world model for multi-agent cooperation by factorizing the naturally composable joint actions of multiple agents and compositionally generating the video conditioned on the world state. By leveraging this compositional world model, in combination with Vision Language Models to infer the actions of other agents, we can use a tree search procedure to integrate these modules and facilitate online cooperative planning. We evaluate our methods on three challenging benchmarks with 2-4 agents. The results show our compositional world model is effective and the framework enables the embodied agents to cooperate efficiently with different agents across various tasks and an arbitrary number of agents, showing the promising future of our proposed methods. More videos can be found at https://umass-embodied-agi.github.io/COMBO

Safe learning is central to AI-enabled robots where a single failure may lead to catastrophic results. Existing safe learning methods are not scalable, inefficient and hard to train, and tend to generate unstable signals under noisy inputs that are challenging to be deployed for robots. To address these challenges, we propose Adaptive explicit-Barrier Net (ABNet) in which barriers explicitly show up in the closed-form model that guarantees safety. The ABNet has the potential to incrementally scale toward larger safe foundation models. Each head of ABNet could learn safe control policies from different features and focuses on specific part of the observation. In this way, we do not need to directly construct a large model for complex tasks, which significantly facilitates the training of the model while ensuring its stable output. Most importantly, we can still formally prove the safety guarantees of the ABNet. We demonstrate the efficiency and strength of ABNet in 2D robot obstacle avoidance, safe robot manipulation, and vision-based end-to-end autonomous driving, with results showing much better robustness and guarantees over existing models.

World models aim to learn action-controlled future prediction and have proven essential for the development of intelligent agents. However, most existing world models rely heavily on substantial action-labeled data and costly training, making it challenging to adapt to novel environments with heterogeneous actions through limited interactions. This limitation can hinder their applicability across broader domains. To overcome this limitation, we propose AdaWorld, an innovative world model learning approach that enables efficient adaptation. The key idea is to incorporate action information during the pretraining of world models. This is achieved by extracting latent actions from videos in a self-supervised manner, capturing the most critical transitions between frames. We then develop an autoregressive world model that conditions on these latent actions. This learning paradigm enables highly adaptable world models, facilitating efficient transfer and learning of new actions even with limited interactions and finetuning. Our comprehensive experiments across multiple environments demonstrate that AdaWorld achieves superior performance in both simulation quality and visual planning.

Large Language Models (LLMs) are increasingly used in applications requiring long context lengths, but the key-value (KV) cache often becomes a memory bottleneck on GPUs as context grows. To address this, we propose Commutative Vector Quantization (CommVQ) to significantly reduce memory usage for long-context LLM inference. We first introduce additive quantization with a lightweight encoder and codebook to compress the KV cache, which can be decoded via simple matrix multiplication. To further reduce computational costs during decoding, we design the codebook to be commutative with Rotary Position Embedding (RoPE) and train it using an Expectation-Maximization (EM) algorithm. This enables efficient integration of decoding into the self-attention mechanism. Our approach achieves high accuracy with additive quantization and low overhead via the RoPE-commutative codebook. Experiments on long-context benchmarks and GSM8K show that our method reduces FP16 KV cache size by 87.5% with 2-bit quantization, while outperforming state-of-the-art KV cache quantization methods. Notably, it enables 1-bit KV cache quantization with minimal accuracy loss, allowing a LLaMA-3.1 8B model to run with a 128K context length on a single RTX 4090 GPU. The source code is available at: https://github.com/UMass-Embodied-AGI/CommVQ.

Large language models (LLMs) have demonstrated remarkable reasoning capabilities across diverse domains. Recent studies have shown that increasing test-time computation enhances LLMs' reasoning capabilities. This typically involves extensive sampling at inference time guided by an external LLM verifier, resulting in a two-player system. Despite external guidance, the effectiveness of this system demonstrates the potential of a single LLM to tackle complex tasks. Thus, we pose a new research problem: *Can we internalize the searching capabilities to fundamentally enhance the reasoning abilities of a single LLM?* This work explores an orthogonal direction focusing on post-training LLMs for autoregressive searching (*i.e.,* an extended reasoning process with self-reflection and self-exploration of new strategies). To achieve this, we propose the Chain-of-Action-Thought (COAT) reasoning and a two-stage training paradigm: 1) a small-scale format tuning stage to internalize the COAT reasoning format and 2) a large-scale self-improvement stage leveraging reinforcement learning. Our approach results in Satori, a 7B LLM trained on open-source models and data. Extensive empirical evaluations demonstrate that Satori achieves state-of-the-art performance on mathematical reasoning benchmarks while exhibits strong generalization to out-of-domain tasks. Code, data, and models are fully open-sourced.

Robotic assembly remains a significant challenge due to complexities in visual perception, functional grasping, contact-rich manipulation, and performing high-precision tasks. Simulation-based learning and sim-to-real transfer have led to recent success in solving assembly tasks in the presence of object pose variation, perception noise, and control error; however, the development of a generalist (i.e., multi-task) agent for a broad range of assembly tasks has been limited by the need to manually curate assembly assets, which greatly constrains the number and diversity of assembly problems that can be used for policy learning. Inspired by recent success of using generative AI to scale up robot learning, we propose Match-Maker, a pipeline to automatically generate diverse, simulation-compatible assembly asset pairs to facilitate learning assembly skills. Specifically, MatchMaker can 1) take a simulation-incompatible, interpenetrating asset pair as input, and automatically convert it into a simulation-compatible, interpenetration-free pair, 2) take an arbitrary single asset as input, and generate a geometrically-mating asset to create an asset pair, 3) automatically erode contact surfaces from (1) or (2) according to a user-specified clearance parameter to generate realistic parts. We demonstrate that data generated by MatchMaker outperforms previous work in terms of diversity and effectiveness for downstream assembly skill learning. Project page: https://wangyian-me.github.io/MatchMaker/.

In this work, we present TalkCuts, a large-scale dataset designed to facilitate the study of multi-shot human speech video generation. Unlike existing datasets that focus on single-shot, static viewpoints, TalkCuts offers 164k clips totaling over 500 hours of high-quality 1080P human speech videos with diverse camera shots, including close-up, half-body, and full-body views. The dataset includes detailed textual descriptions, 2D keypoints and 3D SMPL-X motion annotations, covering over 10k identities, enabling multimodal learning and evaluation. As a first attempt to showcase the value of the dataset, we present Orator, an LLM-guided multi-modal generation framework as a simple baseline, where the language model functions as a multi-faceted director, orchestrating detailed specifications for camera transitions, speaker gesticulations, and vocal modulation. This architecture enables the synthesis of coherent long-form videos through our integrated multi-modal video generation module. Extensive experiments in both pose-guided and audio-driven settings show that training on TalkCuts significantly enhances the cinematographic coherence and visual appeal of generated multi-shot speech videos. We believe TalkCuts provides a strong foundation for future work in controllable, multi-shot speech video generation and broader multimodal learning.

In this paper, we introduce LangSplatV2, which achieves high-dimensional feature splatting at 476.2 FPS and 3D open-vocabulary text querying at 384.6 FPS for high-resolution images, providing a 42 × speedup and a 47 × boost over LangSplat respectively, along with improved query accuracy. LangSplat employs Gaussian Splatting to embed 2D CLIP language features into 3D, significantly enhancing speed and learning a precise 3D language field with SAM semantics. Such advancements in 3D language fields are crucial for applications that require language interaction within complex scenes. However, LangSplat does not yet achieve real- time performance (8.2 FPS), even with advanced A100 GPUs, severely limiting its broader application. In this paper, we first conduct a detailed time analysis of LangSplat, identifying the heavyweight decoder as the primary speed bottleneck. Our solution, LangSplatV2 assumes that each Gaussian acts as a sparse code within a global dictionary, leading to the learning of a 3D sparse coefficient field that entirely eliminates the need for a heavyweight decoder. By leveraging this sparsity, we further propose an efficient sparse coefficient splatting method with CUDA optimization, rendering high-dimensional feature maps at high quality while incurring only the time cost of splatting an ultra-low-dimensional feature. Our experimental results demonstrate that LangSplatV2 not only achieves better or competitive query accuracy but is also significantly faster. Codes and demos are available at our project page: https://langsplat-v2.github.io.

We introduce CameraBench, a large-scale dataset and benchmark designed to assess and improve camera motion understanding. CameraBench consists of ~3,000 diverse internet videos, annotated by experts through a rigorous multi-stage quality control process. One of our core contributions is a taxonomy or "language" of camera motion primitives, designed in collaboration with cinematographers. We find, for example, that some motions like "follow" (or tracking) require understanding scene content like moving subjects. We conduct a large-scale human study to quantify human performance, revealing that domain expertise and tutorial-based training can significantly enhance accuracy. For example, a novice may confuse zoom-in (a change of intrinsics) with translating forward (a change of extrinsics), but can be trained to differentiate the two. Using CameraBench, we evaluate Structure-from-Motion (SfM) and Video-Language Models (VLMs), finding that SfM models struggle to capture semantic primitives that depend on scene content, while generative VLMs struggle to capture geometric primitives that require precise estimation of trajectories. We then fine-tune a generative VLM on CameraBench to achieve the best of both worlds and showcase its applications, including motion-augmented captioning, video question answering, and video-text retrieval. We hope our taxonomy, benchmark, and tutorials will drive future efforts towards the ultimate goal of understanding camera motions in any video.

Endowing robots with tool design abilities is critical for enabling them to solve complex manipulation tasks that would otherwise be intractable. While recent generative frameworks can automatically synthesize task settings—such as 3D scenes and reward functions—they have not yet addressed the challenge of tool-use scenarios. Simply retrieving human-designed tools might not be ideal since many tools (e.g., a rolling pin) are difficult for robotic manipulators to handle. Furthermore, existing tool design approaches either rely on predefined templates with limited parameter tuning or apply generic 3D generation methods that are not optimized for tool creation. To address these limitations, we propose **RobotSmith**, an automated pipeline that leverages the implicit physical knowledge embedded in vision-language models (VLMs) alongside the more accurate physics provided by physics simulations to design and use tools for robotic manipulation. Our system (1) iteratively proposes tool designs using collaborative VLM agents, (2) generates low-level robot trajectories for tool use, and (3) jointly optimizes tool geometry and usage for task performance. We evaluate our approach across a wide range of manipulation tasks involving rigid, deformable, and fluid objects. Experiments show that our method consistently outperforms strong baselines in both task success rate and overall performance. Notably, our approach achieves a 50.0\% average success rate, significantly surpassing other baselines such as 3D generation (21.4\%) and tool retrieval (11.1\%). Finally, we deploy our system in real-world settings, demonstrating that the generated tools and their usage plans transfer effectively to physical execution, validating the practicality and generalization capabilities of our approach.

Spatial reasoning in 3D space is central to human cognition and indispensable for embodied tasks such as navigation and manipulation. However, state-of-the-art vision–language models (VLMs) struggle frequently with tasks as simple as anticipating how a scene will look after an egocentric motion: they perceive 2D images but lack an internal model of 3D dynamics. We therefore propose SpatialNavigator, a test-time scaling framework that grants a VLM with this missing capability by coupling it to a controllable world model based on video diffusion. The VLM iteratively sketches a concise camera trajectory, while the world model synthesizes the corresponding view at each step. The VLM then reasons over this multi-view evidence gathered during the interactive exploration. Without any fine-tuning, our SpatialNavigator achieves an average 7.7\% performance boost on the representative spatial reasoning benchmark SAT, showing that pairing VLMs with world models for test-time scaling offers a simple, plug-and-play route to robust 3D reasoning. Meanwhile, our method also improves upon the test-time inference VLMs trained through reinforcement learning, which demonstrates the potential of our method that utilizes world models for test-time scaling.

The ability to simulate the effects of future actions on the world is a crucial ability of intelligent embodied agents, enabling agents to anticipate the effects of their actions and make plans accordingly. While a large body of existing work has explored how to construct such world models using video models, they are often myopic in nature, without any memory of a scene not captured by currently observed images, preventing agents from making consistent long-horizon plans in complex environments where many parts of the scene are partially observed. We introduce a new persistent embodied world model with an explicit memory of previously generated content, enabling much more consistent long-horizon simulation. During generation time, our video diffusion model predicts RGB-D video of the future observations of the agent. This generation is then aggregated into a persistent 3D map of the environment. By conditioning the video model on this 3D spatial map, we illustrate how this enables video world models to faithfully simulate both seen and unseen parts of the world. Finally, we illustrate the efficacy of such a world model in downstream embodied applications, enabling effective planning and policy learning.

Understanding and reasoning about objects' physical properties in the natural world is a fundamental challenge in artificial intelligence. While some properties like colors and shapes can be directly observed, others, such as mass and electric charge, are hidden from the objects' visual appearance. This paper addresses the unique challenge of inferring these hidden physical properties from objects' motion and interactions and predicting corresponding dynamics based on the inferred physical properties. We first introduce the Compositional Physical Reasoning (ComPhy) dataset. For a given set of objects, ComPhy includes limited videos of them moving and interacting under different initial conditions. The model is evaluated based on its capability to unravel the compositional hidden properties, such as mass and charge, and use this knowledge to answer a set of questions. Besides the synthetic videos from simulators, we also collect a real-world dataset to show further test physical reasoning abilities of different models. We evaluate state-of-the-art video reasoning models on ComPhy and reveal their limited ability to capture these hidden properties, which leads to inferior performance. We also propose a novel neuro-symbolic framework, Physical Concept Reasoner (PCR), that learns and reasons about both visible and hidden physical properties from question answering. Leveraging an object-centric representation, PCR utilizes videos and the associated natural language to infer objects' physical properties without dense object annotations. Furthermore, It incorporates property-aware graph networks to approximate the dynamic interactions among objects. PCR also employs a semantic parser to convert questions into semantic programs, and a program executor to execute the programs based on the learned physical properties and dynamics. After training, PCR demonstrates remarkable capabilities. It can detect and associate objects across frames, ground visible and hidden physical properties, make future and counterfactual predictions, and utilize these extracted representations to answer challenging questions. We hope the proposed ComPhy dataset and the PCR model present a promising step towards more comprehensive physical reasoning in AI systems.

In multiagent systems, learning optimal behavior policies for individual agents remains a challenging yet crucial task. While recent research has made strides in this area, the issue of when agents should maintain consistent behaviors with one another is still not adequately addressed. This article proposes a novel approach to enable agents to autonomously decide whether their behaviors should align with those of their peers by leveraging intrinsic rewards to optimize their policies. We define behavior consistency as the divergence between the actions taken by two agents given the same observations. To encourage agents to be aware of each other's behaviors, we propose dynamic consistency-based intrinsic reward (DCIR), which guides agents in determining when to synchronize their behaviors. In addition, we introduce a dynamic scaling network (DSN) that provides learnable scaling factors at each time step, enabling agents to dynamically decide the extent of rewarding consistent behavior. Our method is evaluated on environments including Multiagent Particle, Google Research Football, and StarCraft II Micromanagement. Experimental results demonstrate its effectiveness in learning optimal policies.

Humans possess the ability to master a wide range of motor skills, enabling them to quickly and flexibly adapt to the surrounding environment. Despite recent progress in replicating such versatile human motor skills, existing research often oversimplifies or inadequately captures the complex interplay between human body movements and highly dynamic environments, such as interactions with fluids. In this paper, we present a world model for Character-Fluid Coupling (CFC) for simulating human-fluid interactions via two-way coupling. We introduce a two-level world model which consists of a Physics-Informed Neural Network (PINN)-based model for fluid dynamics and a character world model capturing body dynamics under various external forces. This two-level world model adeptly predicts the dynamics of fluid and its influence on rigid bodies via force prediction, sidestepping the computational burden of fluid simulation and providing policy gradients for efficient policy training. Once trained, our system can control characters to complete high-level tasks while adaptively responding to environmental changes. We also present that the fluid initiates emergent behaviors of the characters, enhancing motion diversity and interactivity. Extensive experiments underscore the effectiveness of CFC, demonstrating its ability to produce high-quality, realistic human-fluid interaction animations.

Knowledge-based visual reasoning remains a daunting task since it not only requires machines to interpret the concepts and relationships from visual scenes but also associate them with external world knowledge to conduct a chain of reasoning on open-world questions. Previous works, however, treat visual perception and language-based reasoning as two independent modules, failing to attend to both modules throughout all stages of reasoning. To this end, we propose Visual Chain-of-thought Prompting (VCTP) for knowledge-based reasoning, which involves the interaction between visual content and natural language in an iterative step-by-step reasoning manner. VCTP contains three stages, see, think, and confirm. The see stage scans the image and grounds the visual concept candidates with a visual perception model. The think stage adopts a pre-trained large language model (LLM) to attend to key visual concepts from natural language questions adaptively. It then transforms key visual context into text context for prompting with a visual captioning model, and adopts the LLM to generate the answer. The confirm stage further uses the LLM to generate the supporting rationale to the answer, which is then passed through a cross-modality classifier to verify that it’s consistent with the visual context. We iterate through the think-confirm stages to ensure the verified rationale is consistent with the answer. We conduct experiments on a range of knowledge-based visual reasoning datasets. We found our VCTP enjoys several benefits, 1). it achieves better performance than the previous few-shot learning baselines; 2). it enjoys the total transparency and trustworthiness of the whole reasoning process by providing rationales for each reasoning step; 3). it is computation-efficient compared with other fine-tuning baselines. Our code is available at https://github.com/UMass-Foundation-Model/VisualCoT.git

Human beings possess the capability to multiply a mélange of multisensory cues while actively exploring and interacting with the 3D world. Current multi-modal large language models, however, passively absorb sensory data as inputs, lacking the capacity to actively interact with the objects in the 3D environment and dynamically collect their multisensory information. To usher in the study of this area, we propose MultiPLY, a multisensory embodied large language model that could incorporate multisensory interactive data, including visual, audio, tactile, and thermal information into large language models, thereby establishing the correlation among words, actions, and percepts. To this end, we first collect Multisensory Universe, a large-scale multisensory interaction dataset comprising 500k data by deploying an LLM-powered embodied agent to engage with the 3D environment. To perform instruction tuning with pre-trained LLM on such generated data, we first encode the 3D scene as abstracted object-centric representations, and then introduce action tokens denoting that the embodied agent takes certain actions within the environment, as well as state tokens that represent the multisensory state observations of the agent at each time step. In the inference time, MultiPLY could generate action tokens, instructing the agent to take the action in the environment and obtain the next multisensory state observation. The observation is then appended back to the LLM via state tokens to generate subsequent text or action tokens. We demonstrate that MultiPLY out-performs baselines by a large margin through a diverse set of embodied tasks involving object retrieval, tool use, mul-tisensory captioning, and task decomposition.

We introduce Open3DIS, a novel solution designed to tackle the problem of Open-Vocabulary Instance Segmentation within 3D scenes. Objects within 3D environments exhibit diverse shapes, scales, and colors, making precise instance-level identification a challenging task. Recent advancements in Open-Vocabulary scene understanding have made significant strides in this area by employing class-agnostic 3D instance proposal networks for object localization and learning queryable features for each 3D mask. While these methods produce high-quality instance proposals, they struggle with identifying small-scale and geometrically ambiguous objects. The key idea of our method is a new module that aggregates 2D instance masks across frames and maps them to geometrically coherent point cloud regions as high-quality object proposals addressing the above limitations. These are then combined with 3D class-agnostic instance proposals to include a wide range of objects in the real world. To validate our approach, we conducted experiments on three prominent datasets, including ScanNet200, S3DIS, and Replica, demonstrating significant performance gains in segmenting objects with diverse categories over the state-of-the-art approaches.

Learning commonsense reasoning from visual contexts and scenes in real-world is a crucial step toward advanced artificial intelligence. However, existing video reasoning benchmarks are still inadequate since they were mainly designed for factual or situated reasoning and rarely involve broader knowledge in the real world. Our work aims to delve deeper into reasoning evaluations, specifically within dynamic, open-world, and structured context knowledge. We propose a new benchmark (SOK-Bench), consisting of 44K questions and 10K situations with instance-level annotations depicted in the videos. The reasoning process is required to understand and apply situated knowledge and general knowledge for problem-solving. To create such a dataset, we propose an automatic and scalable gener-ation method to generate question-answer pairs, knowledge graphs, and rationales by instructing the combinations of LLMs and MLLMs. Concretely, we first extract observable situated entities, relations, and processes from videos for situated knowledge and then extend to open-world knowledge beyond the visible content. The task generation is facilitated through multiple dialogues as iterations and subsequently corrected and refined by our designed self-promptings and demonstrations. With a corpus of both explicit situated facts and implicit commonsense, we generate associated question-answer pairs and reasoning processes, finally followed by manual reviews for quality assurance. We evaluated recent mainstream large vision-language models on the benchmark and found several in-sightful conclusions. For more information, please refer to our benchmark at www.bobbywu.com/SOKBench.

We leverage Large Language Models (LLM) for zero-shot Semantic Audio Visual Navigation (SAVN). Existing methods utilize extensive training demonstrations for rein-forcement learning, yet achieve relatively low success rates and lack generalizability. The intermittent nature of au-ditory signals further poses additional obstacles to infer-ring the goal information. To address this challenge, we present the Reflective and Imaginative Language Agent (RILA). By employing multi-modal models to process sen-sory data, we instruct an LLM-based planner to actively ex-plore the environment. During the exploration, our agent adaptively evaluates and dismisses inaccurate perceptual descriptions. Additionally, we introduce an auxiliary LLM-based assistant to enhance global environmental compre-hension by mapping room layouts and providing strategic insights. Through comprehensive experiments and analy-sis, we show that our method outperforms relevant base-lines without training demonstrations from the environment and complementary semantic information.

Recent works have shown that Large Language Models (LLMs) could empower traditional neuro-symbolic models via programming capabilities to translate languages into module descriptions, thus achieving strong visual reasoning results while maintaining the model’s transparency and efficiency. However, these models usually exhaustively generate the entire code snippet given each new instance of a task, which is extremely ineffective. On the contrary, human beings gradually acquire knowledge that can be reused and grow into more profound skills for fast generalization to new tasks since we are an infant. Inspired by this, we propose generative neuro-symbolic visual reasoning by growing and reusing modules. Specifically, our model consists of three unique stages, module initialization, module generation, and module execution. First, given a vision-language task, we adopt LLMs to examine whether we could reuse and grow over established modules to handle this new task. If not, we initialize a new module needed by the task and specify the inputs and outputs of this new module. After that, the new module is created by querying LLMs to generate corresponding code snippets that match the requirements. In order to get a better sense of the new module’s ability, we treat few-shot training examples as test cases to see if our new module could pass these cases. If yes, the new module is added to the module library for future reuse. Finally, we evaluate the performance of our model on the testing set by executing the parsed programs with the newly made visual modules to get the results. We find the proposed GENOME model possesses several advantages. First, it performs competitively on standard tasks like visual question answering and referring expression comprehension; Second, the visual modules learned from one task can be seamlessly transferred to new tasks; Last but not least, it is able to adapt to new visual reasoning tasks by observing a few training examples and reusing modules.

A remarkable ability of human beings resides in compositional reasoning, i.e., the capacity to make "infinite use of finite means". However, current large vision-language foundation models (VLMs) fall short of such compositional abilities due to their ``bag-of-words" behaviors and inability to construct words that correctly represent visual entities and the relations among the entities. To this end, we propose CoVLM, which can guide the LLM to explicitly compose visual entities and relationships among the text and dynamically communicate with the vision encoder and detection network to achieve vision-language communicative decoding. Specifically, we first devise a set of novel communication tokens for the LLM, for dynamic communication between the visual detection system and the language system. A communication token is generated by the LLM following a visual entity or a relation, to inform the detection network to propose regions that are relevant to the sentence generated so far. The proposed regions-of-interests (ROIs) are then fed back into the LLM for better language generation contingent on the relevant regions. The LLM is thus able to compose the visual entities and relationships through the communication tokens. The vision-to-language and language-to-vision communication are iteratively performed until the entire sentence is generated. Our framework seamlessly bridges the gap between visual perception and LLMs and outperforms previous VLMs by a large margin on compositional reasoning benchmarks (e.g., ~20% in HICO-DET mAP, ~14% in Cola top-1 accuracy, and ~3% on ARO top-1 accuracy). We also achieve state-of-the-art performances on traditional vision-language tasks such as referring expression comprehension and visual question answering.

We introduce DIFFTACTILE, a physics-based differentiable tactile simulation system designed to enhance robotic manipulation with dense and physically accurate tactile feedback. In contrast to prior tactile simulators which primarily focus on manipulating rigid bodies and often rely on simplified approximations to model stress and deformations of materials in contact, DIFFTACTILE emphasizes physics-based contact modeling with high fidelity, supporting simulations of diverse contact modes and interactions with objects possessing a wide range of material properties. Our system incorporates several key components, including a Finite Element Method (FEM)-based soft body model for simulating the sensing elastomer, a multi-material simulator for modeling diverse object types (such as elastic, elastoplastic, cables) under manipulation, a penalty-based contact model for handling contact dynamics. The differentiable nature of our system facilitates gradient-based optimization for both 1) refining physical properties in simulation using real-world data, hence narrowing the sim-to-real gap and 2) efficient learning of tactile-assisted grasping and contact-rich manipulation skills. Additionally, we introduce a method to infer the optical response of our tactile sensor to contact using an efficient pixel-based neural module. We anticipate that DIFFTACTILE will serve as a useful platform for studying contact-rich manipulations, leveraging the benefits of dense tactile feedback and differentiable physics. Code and supplementary materials are available at the project website https://difftactile.github.io/.

Supervised Fine-Tuning (SFT) on response demonstrations combined with Reinforcement Learning from Human Feedback (RLHF) constitutes a powerful paradigm for aligning LLM-based AI agents. However, a significant limitation of such an approach is its dependency on high-quality human annotations, making its application to intricate tasks challenging due to difficulties in obtaining consistent response demonstrations and in-distribution response preferences. This paper presents a novel approach, namely SALMON, to align base language models with minimal human supervision, using only a small set of human-defined principles, yet achieving superior performance. Central to our approach is an instructable reward model. Trained on synthetic preference data, this model can generate reward scores based on arbitrary human-defined principles. By merely adjusting these principles during the RL training phase, we gain full control over the preferences with the instructable reward model, subsequently influencing the behavior of the RL-trained policy models, and reducing the reliance on the collection of online human preferences. Applying our method to the LLaMA-2-70b base language model, we developed an AI assistant named Dromedary-2. With only 6 exemplars for in-context learning and 31 human-defined principles, Dromedary-2 significantly surpasses the performance of several state-of-the-art AI systems, including LLaMA-2-Chat-70b, on various benchmark datasets. We have open-sourced the code and model weights to encourage further research into aligning LLM-based AI agents with enhanced supervision efficiency, improved controllability, and scalable oversight.

In this work, we aim to teach robots to manipulate various thin-shell materials. Prior works studying thin-shell object manipulation mostly rely on heuristic policies or learn policies from real-world video demonstrations, and only focus on limited material types and tasks (e.g., cloth unfolding). However, these approaches face significant challenges when extended to a wider variety of thin-shell materials and a diverse range of tasks. On the other hand, while virtual simulations are shown to be effective in diverse robot skill learning and evaluation, prior thin-shell simulation environments only support a subset of thin-shell materials, which also limits their supported range of tasks. To fill in this gap, we introduce ThinShellLab - a fully differentiable simulation platform tailored for robotic interactions with diverse thin-shell materials possessing varying material properties, enabling flexible thin-shell manipulation skill learning and evaluation. Building on top of our developed simulation engine, we design a diverse set of manipulation tasks centered around different thin-shell objects. Our experiments suggest that manipulating thin-shell objects presents several unique challenges: 1) thin-shell manipulation relies heavily on frictional forces due to the objects' co-dimensional nature, 2) the materials being manipulated are highly sensitive to minimal variations in interaction actions, and 3) the constant and frequent alteration in contact pairs makes trajectory optimization methods susceptible to local optima, and neither standard reinforcement learning algorithms nor trajectory optimization methods (either gradient-based or gradient-free) are able to solve the tasks alone. To overcome these challenges, we present an optimization scheme that couples sampling-based trajectory optimization and gradient-based optimization, boosting both learning efficiency and converged performance across various proposed tasks. In addition, the differentiable nature of our platform facilitates a smooth sim-to-real transition. By tuning simulation parameters with a minimal set of real-world data, we demonstrate successful deployment of the learned skills to real-robot settings. ThinShellLab will be publicly available. Video demonstration and more information can be found on the project website https://vis-www.cs.umass.edu/ThinShellLab/.

In this work, we address challenging multi-agent cooperation problems with decentralized control, raw sensory observations, costly communication, and multi-objective tasks instantiated in various embodied environments. While previous research either presupposes a cost-free communication channel or relies on a centralized controller with shared observations, we harness the commonsense knowledge, reasoning ability, language comprehension, and text generation prowess of LLMs and seamlessly incorporate them into a cognitive-inspired modular framework that integrates with perception, memory, and execution. Thus building a Cooperative Embodied Language Agent CoELA, who can plan, communicate, and cooperate with others to accomplish long-horizon tasks efficiently. Our experiments on C-WAH and TDW-MAT demonstrate that CoELA driven by GPT-4 can surpass strong planning-based methods and exhibit emergent effective communication. Though current Open LMs like LLAMA-2 still underperform, we fine-tune a CoELA with data collected with our agents and show how they can achieve promising performance. We also conducted a user study for human-agent interaction and discovered that CoELA communicating in natural language can earn more trust and cooperate more effectively with humans. Our research underscores the potential of LLMs for future research in multi-agent cooperation. Videos can be found on the project website https://vis-www.cs.umass.edu/Co-LLM-Agents/.

Recent advances in high-fidelity virtual environments serve as one of the major driving forces for building intelligent embodied agents to perceive, reason and interact with the physical world. Typically, these environments remain unchanged unless agents interact with them. However, in real-world scenarios, agents might also face dynamically changing environments characterized by unexpected events and need to rapidly take action accordingly. To remedy this gap, we propose a new simulated embodied benchmark, called HAZARD, specifically designed to assess the decision-making abilities of embodied agents in dynamic situations. HAZARD consists of three unexpected disaster scenarios, including fire, flood, and wind, and specifically supports the utilization of large language models (LLMs) to assist common sense reasoning and decision-making. This benchmark enables us to evaluate autonomous agents' decision-making capabilities across various pipelines, including reinforcement learning (RL), rule-based, and search-based methods in dynamically changing environments. As a first step toward addressing this challenge using large language models, we further develop an LLM-based agent and perform an in-depth analysis of its promise and challenge of solving these challenging tasks. HAZARD is available at https://vis-www.cs.umass.edu/hazard/.

While self-supervised learning (SSL) in speech has greatly reduced the reliance of speech processing systems on annotated corpora, the success of SSL still hinges on the availability of a large-scale unannotated corpus, which is still often impractical for many low-resource languages or under privacy concerns. Some existing work seeks to alleviate the problem by data augmentation, but most works are confined to introducing perturbations to real speech and do not introduce new variations in speech prosody, speakers, and speech content, which are important for SSL. Motivated by the recent finding that diffusion models have superior capabilities for modeling data distributions, we propose DiffS4L, a pretraining scheme that augments the limited unannotated data with synthetic data with different levels of variations, generated by a diffusion model trained on the limited unannotated data. Finally, an SSL model is pre-trained on the real and the synthetic speech. Our experiments show that DiffS4L can significantly improve the performance of SSL models, such as reducing the WER of the HuBERT pretrained model by 6.26 percentage points in the English ASR task. Notably, we find that the synthetic speech with all levels of variations, i.e. new prosody, new speakers, and even new content (despite the new content being mostly babble), accounts for significant performance improvement. The code is available at github.com/Hertin/DiffS4L.

Large Language Models have recently gained significant attention in scientific discovery for their extensive knowledge and advanced reasoning capabilities. However, they encounter challenges in effectively simulating observational feedback and grounding it with language to propel advancements in physical scientific discovery. Conversely, human scientists undertake scientific discovery by formulating hypotheses, conducting experiments, and revising theories through observational analysis. Inspired by this, we propose to enhance the knowledge-driven, abstract reasoning abilities of LLMs with the computational strength of simulations. We introduce Scientific Generative Agent (SGA), a bilevel optimization framework: LLMs act as knowledgeable and versatile thinkers, proposing scientific hypotheses and reason about discrete components, such as physics equations or molecule structures; meanwhile, simulations function as experimental platforms, providing observational feedback and optimizing via differentiability for continuous parts, such as physical parameters. We conduct extensive experiments to demonstrate our framework’s efficacy in constitutive law discovery and molecular design, unveiling novel solutions that differ from conventional human expectations yet remain coherent upon analysis.

We present RoboGen, a generative robotic agent that automatically learns diverse robotic skills at scale via generative simulation. RoboGen leverages the latest advancements in foundation and generative models. Instead of directly adapting these models to produce policies or low-level actions, we advocate for a generative scheme, which uses these models to automatically generate diversified tasks, scenes, and training supervisions, thereby scaling up robotic skill learning with minimal human supervision. Our approach equips a robotic agent with a self-guided propose-generate-learn cycle: the agent first proposes interesting tasks and skills to develop, and then generates simulation environments by populating pertinent assets with proper spatial configurations. Afterwards, the agent decomposes the proposed task into sub-tasks, selects the optimal learning approach (reinforcement learning, motion planning, or trajectory optimization), generates required training supervision, and then learns policies to acquire the proposed skill. Our fully generative pipeline can be queried repeatedly, producing an endless stream of skill demonstrations associated with diverse tasks and environments.

Recent vision-language-action (VLA) models rely on 2D inputs, lacking integration with the broader realm of the 3D physical world. Furthermore, they perform action prediction by learning a direct mapping from perception to action, neglecting the vast dynamics of the world and the relations between actions and dynamics. In contrast, human beings are endowed with world models that depict imagination about future scenarios to plan action accordingly. To this end, we propose 3D-VLA by introducing a new family of embodied foundation models that seamlessly link 3D perception, reasoning, and action through a generative world model. Specifically, 3D-VLA is built on top of a 3D-based large language model (LLM) and a set of action tokens is introduced to engage with the embodied environment. Furthermore, to inject generation abilities into the model, we train the embodied diffusion models and align them into the LLM for predicting the goal image and point cloud. To train our 3D-VLA, we curate a large-scale 3D embodied instruction dataset by extracting vast 3D-related information from existing robotics datasets. Our experiments on held-in datasets demonstrate that 3D-VLA significantly improves the reasoning, multimodality generation and planning capabilities in embodied environments, showcasing its potential in real-world applications.

We introduce the Continuum Physical Dataset (ContPhy), a novel benchmark for assessing machine physical commonsense. ContPhy complements existing physical reasoning benchmarks by encompassing the inference of diverse physical properties, such as mass and density, across various scenarios and predicting corresponding dynamics. We evaluated a range of AI models and found that they still struggle to achieve satisfactory performance on ContPhy, which shows that current AI models still lack physical commonsense for the continuum, especially soft-bodies, and illustrates the value of the proposed dataset. We also introduce an oracle model (ContPRO) that marries the particle-based physical dynamic models with the recent large language models, which enjoy the advantages of both models, precise dynamic predictions, and interpretable reasoning. ContPhy aims to spur progress in perception and reasoning within diverse physical settings, narrowing the divide between human and machine intelligence in understanding the physical world.

Text-to-video models have demonstrated substantial potential in robotic decision-making, enabling the imagination of realistic plans of future actions as well as accurate environment simulation. However, one major issue in such models is generalization – models are limited to synthesizing videos subject to language instructions similar to those seen at training time. This is heavily limiting in decision-making, where we seek a powerful world model to synthesize plans of unseen combinations of objects and actions in order to solve previously unseen tasks in new environments. To resolve this issue, we introduce RoboDreamer, an innovative approach for learning a compositional world model by factorizing the video generation. We leverage the natural compositionality of language to parse instructions into a set of lower-level primitives, which we condition a set of models on to generate videos. We illustrate how this factorization naturally enables compositional generalization, by allowing us to formulate a new natural language instruction as a combination of previously seen components. We further show how such a factorization enables us to add additional multimodal goals, allowing us to specify a video we wish to generate given both natural language instructions and a goal image. Our approach can successfully synthesize video plans on unseen goals in the RT-X, enables successful robot execution in simulation, and substantially outperforms monolithic baseline approaches to video generation.

For robots to perform a wide variety of tasks, they require a 3D representation of the world that is semantically rich, yet compact and efficient for task-driven perception and planning. Recent approaches have attempted to leverage features from large vision-language models to encode semantics in 3D representations. However, these approaches tend to produce maps with per-point feature vectors, which do not scale well in larger environments, nor do they contain semantic spatial relationships between entities in the environment, which are useful for downstream planning. In this work, we propose ConceptGraphs, an open-vocabulary graph-structured representation for 3D scenes. ConceptGraphs is built by leveraging 2D foundation models and fusing their output to 3D by multi-view association. The resulting representations generalize to novel semantic classes, without the need to collect large 3D datasets or finetune models. We demonstrate the utility of this representation through a number of downstream planning tasks that are specified through abstract (language) prompts and require complex reasoning over spatial and semantic concepts. To explore the full scope of our experiments and results, we encourage readers to visit our project webpage.

We study the problem of multimodal physical scene understanding, where an embodied agent needs to find fallen objects by inferring object properties, direction, and distance of an impact sound source. Previous works adopt feed-forward neural networks to directly regress the variables from sound, leading to poor generalization and domain adaptation issues. In this paper, we illustrate that learning a disentangled model of acoustic formation, referred to as disentangled acoustic field (DAF), to capture the sound generation and propagation process, enables the embodied agent to construct a spatial uncertainty map over where the objects may have fallen. We demonstrate that our analysis-by-synthesis framework can jointly infer sound properties by explicitly decomposing and factorizing the latent space of the disentangled model. We further show that the spatial uncertainty map can significantly improve the success rate for the localization of fallen objects by proposing multiple plausible exploration locations.

Social relation reasoning aims to identify relation categories such as friends, spouses, and colleagues from images. While current methods adopt the paradigm of training a dedicated network end-to-end using labeled image data, they are limited in terms of generalizability and interpretability. To address these issues, we first present a simple yet well-crafted framework named SocialGPT, which combines the perception capability of Vision Foundation Models (VFMs) and the reasoning capability of Large Language Models (LLMs) within a modular framework, providing a strong baseline for social relation recognition. Specifically, we instruct VFMs to translate image content into a textual social story, and then utilize LLMs for text-based reasoning. SocialGPT introduces systematic design principles to adapt VFMs and LLMs separately and bridge their gaps. Without additional model training, it achieves competitive zero-shot results on two databases while offering interpretable answers, as LLMs can generate language-based explanations for the decisions. The manual prompt design process for LLMs at the reasoning phase is tedious and an automated prompt optimization method is desired. As we essentially convert a visual classification task into a generative task of LLMs, automatic prompt optimization encounters a unique long prompt optimization issue. To address this issue, we further propose the Greedy Segment Prompt Optimization (GSPO), which performs a greedy search by utilizing gradient information at the segment level. Experimental results show that GSPO significantly improves performance, and our method also generalizes to different image styles. The code is available at https://github.com/Mengzibin/SocialGPT.

We introduce Constrained Human-AI Cooperation (CHAIC), an inclusive embodied social intelligence challenge designed to test social perception and cooperation in embodied agents. In CHAIC, the goal is for an embodied agent equipped with egocentric observations to assist a human who may be operating under physical constraints—e.g., unable to reach high places or confined to a wheelchair—in performing common household or outdoor tasks as efficiently as possible. To achieve this, a successful helper must: (1) infer the human's intents and constraints by following the human and observing their behaviors (social perception), and (2) make a cooperative plan tailored to the human partner to solve the task as quickly as possible, working together as a team (cooperative planning). To benchmark this challenge, we create four new agents with real physical constraints and eight long-horizon tasks featuring both indoor and outdoor scenes with various constraints, emergency events, and potential risks. We benchmark planning- and learning-based baselines on the challenge and introduce a new method that leverages large language models and behavior modeling. Empirical evaluations demonstrate the effectiveness of our benchmark in enabling systematic assessment of key aspects of machine social intelligence. Our benchmark and code are publicly available at https://github.com/UMass-Foundation-Model/CHAIC.

We present a computational framework that transforms single images into 3D physical objects. The visual geometry of a physical object in an image is determined by three orthogonal attributes: mechanical properties, external forces, and rest-shape geometry. Existing single-view 3D reconstruction methods often overlook this underlying composition, presuming rigidity or neglecting external forces. Consequently, the reconstructed objects fail to withstand real-world physical forces, resulting in instability or undesirable deformation -- diverging from their intended designs as depicted in the image. Our optimization framework addresses this by embedding physical compatibility into the reconstruction process. We explicitly decompose the three physical attributes and link them through static equilibrium, which serves as a hard constraint, ensuring that the optimized physical shapes exhibit desired physical behaviors. Evaluations on a dataset collected from Objaverse demonstrate that our framework consistently enhances the physical realism of 3D models over existing methods. The utility of our framework extends to practical applications in dynamic simulations and 3D printing, where adherence to physical compatibility is paramount.

Compositional Reasoning (CR) entails grasping the significance of attributes, relations, and word order. Recent Vision-Language Models (VLMs), comprising a visual encoder and a Large Language Model (LLM) decoder, have demonstrated remarkable proficiency in such reasoning tasks. This prompts a crucial question: have VLMs effectively tackled the CR challenge? We conjecture that existing CR benchmarks may not adequately push the boundaries of modern VLMs due to the reliance on an LLM only negative text generation pipeline. Consequently, the negatives produced either appear as outliers from the natural language distribution learned by VLMs' LLM decoders or as improbable within the corresponding image context. To address these limitations, we introduce ConMe\footnote{ConMe is an abbreviation for Confuse Me.} -- a compositional reasoning benchmark and a novel data generation pipeline leveraging VLMs to produce `hard CR Q&A'. Through a new concept of VLMs conversing with each other to collaboratively expose their weaknesses, our pipeline autonomously generates, evaluates, and selects challenging compositional reasoning questions, establishing a robust CR benchmark, also subsequently validated manually. Our benchmark provokes a noteworthy, up to 33%, decrease in CR performance compared to preceding benchmarks, reinstating the CR challenge even for state-of-the-art VLMs.

Current AI alignment methodologies rely on human-provided demonstrations or judgments, and the learned capabilities of AI systems would be upper-bounded by human capabilities as a result. This raises a challenging research question: How can we keep improving the systems when their capabilities have surpassed the levels of humans? This paper answers this question in the context of tackling hard reasoning tasks (e.g., level 4-5 MATH problems) via learning from human annotations on easier tasks (e.g., level 1-3 MATH problems), which we term as easy-to-hard generalization. Our key insight is that an evaluator (reward model) trained on supervisions for easier tasks can be effectively used for scoring candidate solutions of harder tasks and hence facilitating easy-to-hard generalization over different levels of tasks. Based on this insight, we propose a novel approach to scalable alignment, which firstly trains the (process-supervised) reward models on easy problems (e.g., level 1-3), and then uses them to evaluate the performance of policy models on hard problems. We show that such easy-to-hard generalization from evaluators can enable easy-to-hard generalizations in generators either through re-ranking or reinforcement learning (RL). Notably, our process-supervised 7b RL model and 34b model (reranking@1024) achieves an accuracy of 34.0% and 52.5% on MATH500, respectively, despite only using human supervision on easy problems. Our approach suggests a promising path toward AI systems that advance beyond the frontier of human supervision.

Creating large-scale interactive 3D environments is essential for the development of Robotics and Embodied AI research. However, generating diverse embodied environments with realistic detail and considerable complexity remains a significant challenge. Current methods, including manual design, procedural generation, diffusion-based scene generation, and large language model (LLM) guided scene design, are hindered by limitations such as excessive human effort, reliance on predefined rules or training datasets, and limited 3D spatial reasoning ability. Since pre-trained 2D image generative models better capture scene and object configuration than LLMs, we address these challenges by introducing $\textit{Architect}$, a generative framework that creates complex and realistic 3D embodied environments leveraging diffusion-based 2D image inpainting. In detail, we utilize foundation visual perception models to obtain each generated object from the image and leverage pre-trained depth estimation models to lift the generated 2D image to 3D space. While there are still challenges that the camera parameters and scale of depth are still absent in the generated image, we address those problems by ''controlling'' the diffusion model by $\textit{hierarchical inpainting}$. Specifically, having access to ground-truth depth and camera parameters in simulation, we first render a photo-realistic image of only the background. Then, we inpaint the foreground in this image, passing the geometric cues to the inpainting model in the background, which informs the camera parameters. This process effectively controls the camera parameters and depth scale for the generated image, facilitating the back-projection from 2D image to 3D point clouds. Our pipeline is further extended to a hierarchical and iterative inpainting process to continuously generate the placement of large furniture and small objects to enrich the scene. This iterative structure brings the flexibility for our method to generate or refine scenes from various starting points, such as text, floor plans, or pre-arranged environments. Experimental results demonstrate that $\textit{Architect}$ outperforms existing methods in producing realistic and complex environments, making it highly suitable for Embodied AI and robotics applications.

Optimization in multi-task learning (MTL) is more challenging than single-task learning (STL), as the gradient from different tasks can be contradictory. When tasks are related, it can be beneficial to share some parameters among them (cooperation). However, some tasks require additional parameters with expertise in a specific type of data or discrimination (specialization). To address the MTL challenge, we propose Mod-Squad, a new model that is Modularized into groups of experts (a ‘Squad’). This structure allows us to formalize cooperation and specialization as the process of matching experts and tasks. We optimize this matching process during the training of a single model. Specifically, we incorporate mixture of experts (MoE) layers into a transformer model, with a new loss that incorporates the mutual dependence between tasks and experts. As a result, only a small set of experts are activated for each task. This prevents the sharing of the entire backbone model between all tasks, which strengthens the model, especially when the training set size and the number of tasks scale up. More interestingly, for each task, we can extract the small set of experts as a standalone model that maintains the same performance as the large model. Extensive experiments on the Taskonomy dataset with 13 vision tasks and the PASCAL-Context dataset with 5 vision tasks show the superiority of our approach. The project page can be accessed at https://vis-www.cs.umass.edu/Mod-Squad.

Humans possess a versatile mechanism for extracting structured representations of our visual world. When looking at an image, we can decompose the scene into entities and their parts as well as obtain the dependencies between them. To mimic such capability, we propose Visual Dependency Transformers (DependencyViT)11https://github.com/dingmyu/DependencyViT that can induce visual dependencies without any labels. We achieve that with a novel neural operator called reversed attention that can naturally capture long-range visual dependencies between image patches. Specifically, we formulate it as a dependency graph where a child token in reversed attention is trained to attend to its parent tokens and send information following a normalized probability distribution rather than gathering information in conventional self-attention. With such a design, hierarchies naturally emerge from reversed attention layers, and a dependency tree is progressively induced from leaf nodes to the root node unsupervisedly. DependencyViT offers several appealing benefits. (i) Entities and their parts in an image are represented by different subtrees, enabling part partitioning from dependencies; (ii) Dynamic visual pooling is made possible. The leaf nodes which rarely send messages can be pruned without hindering the model performance, based on which we propose the lightweight DependencyViT-Lite to reduce the computational and memory footprints; (iii) DependencyViT works well on both self- and weakly-supervised pretraining paradigms on ImageNet, and demonstrates its effectiveness on 8 datasets and 5 tasks, such as unsupervised part and saliency segmentation, recognition, and detection.

Humans are able to accurately reason in 3D by gathering multi-view observations of the surrounding world. Inspired by this insight, we introduce a new large-scale benchmark for 3D multi-view visual question answering (3DMV-VQA). This dataset is collected by an embodied agent actively moving and capturing RGB images in an environment using the Habitat simulator. In total, it consists of approximately 5k scenes, 600k images, paired with 50k questions. We evaluate various state-of-the-art models for visual reasoning on our benchmark and find that they all perform poorly. We suggest that a principled approach for 3D reasoning from multi-view images should be to infer a compact 3D representation of the world from the multi-view images, which is further grounded on open-vocabulary semantic concepts, and then to execute reasoning on these 3D representations. As the first step towards this approach, we propose a novel 3D concept learning and reasoning (3D-CLR) framework that seamlessly combines these components via neural fields, 2D pre-trained vision-language models, and neural reasoning operators. Experimental results suggest that our framework outperforms baseline models by a large margin, but the challenge remains largely unsolved. We further perform an in-depth analysis of the challenges and highlight potential future directions..

Answering questions about complex situations in videos requires not only capturing the presence of actors, objects, and their relations but also the evolution of these relationships over time. A situation hyper-graph is a representation that describes situations as scene sub-graphs for video frames and hyper-edges for connected sub-graphs and has been proposed to capture all such information in a compact structured form. In this work, we propose an architecture for Video Question Answering (VQA) that enables answering questions related to video content by predicting situation hyper-graphs, coined Situation Hyper-Graph based Video Question Answering (SHG- VQA). To this end, we train a situation hyper-graph decoder to implicitly identify graph representations with actions and object/human-object relationships from the input video clip. and to use cross-attention between the predicted situation hyper-graphs and the question embedding to predict the correct answer. The proposed method is trained in an end-to-end manner and optimized by a VQA loss with the cross-entropy function and a Hungarian matching loss for the situation graph prediction. The effectiveness of the proposed architecture is extensively evaluated on two challenging benchmarks: AGQA and STAR. Our results show that learning the underlying situation hyper-graphs helps the system to significantly improve its performance for novel challenges of video question-answering tasks11Code will be available at https://github.com/aurooj/SHG-VQA.

Embodied control requires agents to leverage multimodal pre-training to quickly learn how to act in new environments, where video demonstrations contain visual and motion details needed for low-level perception and control, and language instructions support generalization with abstract, symbolic structures. While recent approaches apply contrastive learning to force alignment between the two modalities, we hypothesize better modeling their complementary differences can lead to more holistic representations for downstream adaption. To this end, we propose Emergent Communication for Embodied Control (EC2), a novel scheme to pre-train video-language representations for few-shot embodied control. The key idea is to learn an unsupervised “language” of videos via emergent communication, which bridges the semantics of video details and structures of natural language. We learn embodied representations of video trajectories, emergent language, and natural language using a language model, which is then used to finetune a lightweight policy network for downstream control. Through extensive experiments in Metaworld and Franka Kitchen embodied benchmarks, EC2is shown to consistently outperform previous contrastive learning methods for both videos and texts as task inputs. Further ablations confirm the importance of the emergent language, which is beneficial for both video and language learning, and significantly superior to using pre-trained video captions. We also present a quantitative and qualitative analysis of the emergent language and discuss future directions toward better understanding and leveraging emergent communication in embodied tasks.

Modeling sounds emitted from physical object interactions is critical for immersive perceptual experiences in real and virtual worlds. Traditional methods of impact sound synthesis use physics simulation to obtain a set of physics parameters that could represent and synthesize the sound. However, they require fine details of both the object geometries and impact locations, which are rarely available in the real world and can not be applied to synthesize impact sounds from common videos. On the other hand, existing video-driven deep learning-based approaches could only capture the weak correspondence between visual content and impact sounds since they lack of physics knowledge. In this work, we propose a physics-driven diffusion model that can synthesize high-fidelity impact sound for a silent video clip. In addition to the video content, we propose to use additional physics priors to guide the impact sound synthesis procedure. The physics priors include both physics parameters that are directly estimated from noisy real-world impact sound examples without sophisticated setup and learned residual parameters that interpret the sound environment via neural networks. We further implement a novel diffusion model with specific training and inference strategies to combine physics priors and visual information for impact sound synthesis. Experimental results show that our model outperforms several existing systems in generating realistic impact sounds. Lastly, the physics-based representations are fully interpretable and transparent, thus allowing us to perform sound editing flexibly. We encourage the readers visit our project page11https://sukun1045.github.io/video-physics-sound-diffusion/ to watch demo videos with the audio turned on to experience the result.

How to learn discriminative video representation from unlabeled videos is challenging but crucial for video analysis. The latest attempts seek to learn a representation model by predicting the appearance contents in the masked regions. However, simply masking and recovering appearance contents may not be sufficient to model temporal clues as the appearance contents can be easily reconstructed from a single frame. To overcome this limitation, we present Masked Motion Encoding (MME), a new pretraining paradigm that reconstructs both appearance and motion information to explore temporal clues. In MME, we focus on addressing two critical challenges to improve the representation performance: 1) how to well represent the possible long-term motion across multiple frames; and 2) how to obtain fine-grained temporal clues from sparsely sampled videos. Motivated by the fact that human is able to recognize an action by tracking objects' position changes and shape changes, we propose to reconstruct a motion trajectory that represents these two kinds of change in the masked regions. Besides, given the sparse video input, we enforce the model to reconstruct dense motion trajectories in both spatial and temporal dimensions. Pre-trained with our MME paradigm, the model is able to anticipate long-term and fine-grained motion details. Code is available at https://github.com/XinyuSun/MME.

The Universal Transformer (UT) is a variant of the Transformer that shares parameters across its layers.Empirical evidence shows that UTs have better compositional generalization than Vanilla Transformers (VTs) in formal language tasks.The parameter-sharing also affords it better parameter efficiency than VTs.Despite its many advantages, scaling UT parameters is much more compute and memory intensive than scaling up a VT.This paper proposes the Sparse Universal Transformer (SUT), which leverages Sparse Mixture of Experts (SMoE) and a new stick-breaking-based dynamic halting mechanism to reduce UT's computation complexity while retaining its parameter efficiency and generalization ability.Experiments show that SUT achieves the same performance as strong baseline models while only using half computation and parameters on WMT'14 and strong generalization results on formal language tasks (Logical inference and CFQ).The new halting mechanism also enables around 50% reduction in computation during inference with very little performance decrease on formal language tasks.

High-resolution dense prediction enables many appealing real-world applications, such as computational photography, autonomous driving, etc. However, the vast computational cost makes deploying state-of-the-art high-resolution dense prediction models on hardware devices difficult. This work presents EfficientViT, a new family of high-resolution vision models with novel lightweight multi-scale attention. Unlike prior high-resolution dense prediction models that rely on heavy self-attention, hardware-inefficient large-kernel convolution, or complicated topology structure to obtain good performances, our lightweight multi-scale attention achieves a global receptive field and multi-scale learning (two critical features for high-resolution dense prediction) with only lightweight and hardware-efficient operations. As such, EfficientViT delivers remarkable performance gains over previous state-of-the-art high-resolution dense prediction models with significant speedup on diverse hardware platforms, including mobile CPU, edge GPU, and cloud GPU. Without performance loss on Cityscapes, our EfficientViT provides up to 8.8× and 3.8× GPU latency reduction over SegFormer and SegNeXt, respectively. For super-resolution, EfficientViT provides up to 6.4× speedup over Restormer while providing 0.11dB gain in PSNR.

Vision-and-language navigation (VLN) requires an embodied agent to navigate in realistic 3D environments using natural language instructions. Existing VLN methods suffer from training on small-scale environments or unreasonable path-instruction datasets, limiting the generalization to unseen environments. There are massive house tour videos on YouTube, providing abundant real navigation experiences and layout information. However, these videos have not been explored for VLN before. In this paper, we propose to learn an agent from these videos by creating a large-scale dataset which comprises reasonable path-instruction pairs from house tour videos and pre-training the agent on it. To achieve this, we have to tackle the challenges of automatically constructing path-instruction pairs and exploiting real layout knowledge from raw and unlabeled videos. To address these, we first leverage an entropy-based method to construct the nodes of a path trajectory. Then, we propose an action-aware generator for generating instructions from unlabeled trajectories. Last, we devise a trajectory judgment pretext task to encourage the agent to mine the layout knowledge. Experimental results show that our method achieves state-of-the-art performance on two popular benchmarks (R2R and REVERIE). Code is available at https://github.com/JeremyLinky/YouTube-VLN

Panoptic Scene Graph has recently been proposed for comprehensive scene understanding. However, previous works adopt a fully-supervised learning manner, requiring large amounts of pixel-wise densely-annotated data, which is always tedious and expensive to obtain. To address this limitation, we study a new problem of Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). The key idea is to leverage the large collection of free image-caption data on the Web alone to generate panoptic scene graphs. The problem is very challenging for three constraints: 1) no location priors; 2) no explicit links between visual regions and textual entities; and 3) no predefined concept sets. To tackle this problem, we propose a new framework TextPSG consisting of four modules, i.e., a region grouper, an entity grounder, a segment merger, and a label generator, with several novel techniques. The region grouper first groups image pixels into different segments and the entity grounder then aligns visual segments with language entities based on the textual description of the segment being referred to. The grounding results can thus serve as pseudo labels enabling the segment merger to learn the segment similarity as well as guiding the label generator to learn object semantics and relation predicates, resulting in a fine-grained structured scene understanding. Our framework is effective, significantly outperforming the baselines and achieving strong out-of-distribution robustness. We perform comprehensive ablation studies to corroborate the effectiveness of our design choices and provide an in-depth analysis to highlight future directions. Our code, data, and results are available on our project page: https://vis-www.cs.umass.edu/TextPSG.

We propose a hybrid neural network (NN) and PDE approach for learning generalizable PDE dynamics from motion observations. Many NN approaches learn an end-to-end model that implicitly models both the governing PDE and constitutive models (or material models). Without explicit PDE knowledge, these approaches cannot guarantee physical correctness and have limited generalizability. We argue that the governing PDEs are often well-known and should be explicitly enforced rather than learned. Instead, constitutive models are particularly suitable for learning due to their data-fitting nature. To this end, we introduce a new framework termed "Neural Constitutive Laws" (NCLaw), which utilizes a network architecture that strictly guarantees standard constitutive priors, including rotation equivariance and undeformed state equilibrium. We embed this network inside a differentiable simulation and train the model by minimizing a loss function based on the difference between the simulation and the motion observation. We validate NCLaw on various large-deformation dynamical systems, ranging from solids to fluids. After training on a single motion trajectory, our method generalizes to new geometries, initial/boundary conditions, temporal ranges, and even multi-physics systems. On these extremely out-of-distribution generalization tasks, NCLaw is orders-of-magnitude more accurate than previous NN approaches. Real-world experiments demonstrate our method’s ability to learn constitutive laws from videos.

We propose a new method to ensure neural ordinary differential equations (ODEs) satisfy output specifications by using invariance set propagation. Our approach uses a class of control barrier functions to transform output specifications into constraints on the parameters and inputs of the learning system. This setup allows us to achieve output specification guarantees simply by changing the constrained parameters/inputs both during training and inference. Moreover, we demonstrate that our invariance set propagation through data-controlled neural ODEs not only maintains generalization performance but also creates an additional degree of robustness by enabling causal manipulation of the system’s parameters/inputs. We test our method on a series of representation learning tasks, including modeling physical dynamics and convexity portraits, as well as safe collision avoidance for autonomous vehicles.

We investigate the challenge of parametrizing policies for reinforcement learning (RL) in high-dimensional continuous action spaces. Our objective is to develop a multimodal policy that overcomes limitations inherent in the commonly-used Gaussian parameterization. To achieve this, we propose a principled framework that models the continuous RL policy as a generative model of optimal trajectories. By conditioning the policy on a latent variable, we derive a novel variational bound as the optimization objective, which promotes exploration of the environment. We then present a practical model-based RL method, called Reparameterized Policy Gradient (RPG), which leverages the multimodal policy parameterization and learned world model to achieve strong exploration capabilities and high data efficiency. Empirical results demonstrate that our method can help agents evade local optima in tasks with dense rewards and solve challenging sparse-reward environments by incorporating an object-centric intrinsic reward. Our method consistently outperforms previous approaches across a range of tasks. Code and supplementary materials are available on the project page https://haosulab.github.io/RPG/

Correspondence identification (CoID) is an essential component for collaborative perception in multi-robot systems, such as connected autonomous vehicles. The goal of CoID is to identify the correspondence of objects observed by multiple robots in their own field of view in order for robots to consistently refer to the same objects. CoID is challenging due to perceptual aliasing, object non-covisibility, and noisy sensing. In this paper, we introduce a novel deep masked graph matching approach to enable CoID and address the challenges. Our approach formulates CoID as a graph matching problem and we design a masked neural network to integrate the multimodal visual, spatial, and GPS information to perform CoID. In addition, we design a new technique to explicitly address object non-covisibility caused by occlusion and the vehicle's limited field of view. We evaluate our approach in a variety of street environments using a high-fidelity simulation that integrates the CARLA and SUMO simulators. The experimental results show that our approach outperforms the previous approaches and achieves state-of-the- art CoID performance in connected autonomous driving applications. Our work is available at: https://github.com/gaopeng5/DMGM.git.

On-device learning and efficient fine-tuning enable continuous and privacy-preserving customization (e.g., locally fine-tuning large language models on personalized data). However, existing training frameworks are designed for cloud servers with powerful accelerators (e.g., GPUs, TPUs) and lack the optimizations for learning on the edge, which faces challenges of resource limitations and edge hardware diversity. We introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices. PockEngine supports sparse backpropagation: it prunes the backward graph and sparsely updates the model with measured memory saving and latency reduction while maintaining the model quality. Secondly, PockEngine is compilation first: the entire training graph (including forward, backward and optimization steps) is derived at compile-time, which reduces the runtime overhead and brings opportunities for graph transformations. PockEngine also integrates a rich set of training graph optimizations, thus can further accelerate the training cost, including operator reordering and backend switching. PockEngine supports diverse applications, frontends and hardware backends: it flexibly compiles and tunes models defined in PyTorch/TensorFlow/Jax and deploys binaries to mobile CPU/GPU/DSPs. We evaluated PockEngine on both vision models and large language models. PockEngine achieves up to 15 × speedup over off-the-shelf TensorFlow (Raspberry Pi), 5.6 × memory saving back-propagation (Jetson AGX Orin). Remarkably, PockEngine enables fine-tuning LLaMav2-7B on NVIDIA Jetson AGX Orin at 550 tokens/s, 7.9 × faster than the PyTorch.

Large language models (LLMs) and Vision-Language Models (VLMs) have been proved to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models, and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi-view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs could better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (\textit{e.g.}, the BLEU-1 score surpasses state-of-the-art score by 9\%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs. Our model and data will be publicly available.

Combining gradient-based trajectory optimization with differentiable physics simulation is an efficient technique for solving soft-body manipulation problems. Using a well-crafted optimization objective, the solver can quickly converge onto a valid trajectory. However, writing the appropriate objective functions requires expert knowledge, making it difficult to collect a large set of naturalistic problems from non-expert users. We introduce DiffVL, a method that enables non-expert users to communicate soft-body manipulation tasks -- a combination of vision and natural language, given in multiple stages -- that can be readily leveraged by a differential physics solver. We have developed GUI tools that enable non-expert users to specify 100 tasks inspired by real-life soft-body manipulations from online videos, which we'll make public. We leverage large language models to translate task descriptions into machine-interpretable optimization objectives. The optimization objectives can help differentiable physics solvers to solve these long-horizon multistage tasks that are challenging for previous baselines.

Recent AI-assistant agents, such as ChatGPT, predominantly rely on supervised fine-tuning (SFT) with human annotations and reinforcement learning from human feedback (RLHF) to align the output of large language models (LLMs) with human intentions, ensuring they are helpful, ethical, and reliable. However, this dependence can significantly constrain the true potential of AI-assistant agents due to the high cost of obtaining human supervision and the related issues on quality, reliability, diversity, self-consistency, and undesirable biases. To address these challenges, we propose a novel approach called SELF-ALIGN, which combines principle-driven reasoning and the generative power of LLMs for the self-alignment of AI agents with minimal human supervision. Our approach encompasses four stages: first, we use an LLM to generate synthetic prompts, and a topic-guided method to augment the prompt diversity; second, we use a small set of human-written principles for AI models to follow, and guide the LLM through in-context learning from demonstrations (of principles application) to produce helpful, ethical, and reliable responses to user's queries; third, we fine-tune the original LLM with the high-quality self-aligned responses so that the resulting model can generate desirable responses for each query directly without the principle set and the demonstrations anymore; and finally, we offer a refinement step to address the issues of overly-brief or indirect responses. Applying SELF-ALIGN to the LLaMA-65b base language model, we develop an AI assistant named Dromedary. With fewer than 300 lines of human annotations (including < 200 seed prompts, 16 generic principles, and 5 exemplars for in-context learning). Dromedary significantly surpasses the performance of several state-of-the-art AI systems, including Text-Davinci-003 and Alpaca, on benchmark datasets with various settings.

General physical scene understanding requires more than simply localizing and recognizing objects -- it requires knowledge that objects can have different latent properties (e.g., mass or elasticity), and that those properties affect the outcome of physical events. While there has been great progress in physical and video prediction models in recent years, benchmarks to test their performance typically do not require an understanding that objects have individual physical properties, or at best test only those properties that are directly observable (e.g., size or color). This work proposes a novel dataset and benchmark, termed Physion++, that rigorously evaluates visual physical prediction in artificial systems under circumstances where those predictions rely on accurate estimates of the latent physical properties of objects in the scene. Specifically, we test scenarios where accurate prediction relies on estimates of properties such as mass, friction, elasticity, and deformability, and where the values of those properties can only be inferred by observing how objects move and interact with other objects or fluids. We evaluate the performance of a number of state-of-the-art prediction models that span a variety of levels of learning vs. built-in knowledge, and compare that performance to a set of human predictions. We find that models that have been trained using standard regimes and datasets do not spontaneously learn to make inferences about latent properties, but also that models that encode objectness and physical states tend to make better predictions. However, there is still a huge gap between all models and human performance, and all models' predictions correlate poorly with those made by humans, suggesting that no state-of-the-art model is learning to make physical predictions in a human-like way. These results show that current deep learning models that succeed in some settings nevertheless fail to achieve human-level physical prediction in other cases, especially those where latent property inference is required. Project page: https://dingmyu.github.io/physion_v2/

Nature evolves creatures with a high complexity of morphological and behavioral intelligence, meanwhile computational methods lag in approaching that diversity and efficacy. Co-optimization of artificial creatures' morphology and control in silico shows promise for applications in physical soft robotics and virtual character creation; such approaches, however, require developing new learning algorithms that can reason about function atop pure structure. In this paper, we present DiffuseBot, a physics-augmented diffusion model that generates soft robot morphologies capable of excelling in a wide spectrum of tasks. \name bridges the gap between virtually generated content and physical utility by (i) augmenting the diffusion process with a physical dynamical simulation which provides a certificate of performance, and (ii) introducing a co-design procedure that jointly optimizes physical design and control by leveraging information about physical sensitivities from differentiable simulation. We showcase a range of simulated and fabricated robots along with their capabilities. Check our website: https://diffusebot.github.io/

Diffusion models have risen a promising approach to data-driven planning, and have demonstrated impressive robotic control, reinforcement learning, and video planning performance. Given an effective planner, an important question to consider is replanning -- when given plans should be regenerated due to both action execution error and external environment changes. Direct plan execution, without replanning, is problematic as errors from individual actions rapidly accumulate and environments are partially observable and stochastic. Simultaneously, replanning at each timestep incurs a substantial computational cost, and may prevent successful task execution, as different generated plans prevent consistent progress to any particular goal. In this paper, we explore how we may effectively replan with diffusion models. We propose a principled approach to determine when to replan, based on the diffusion model's estimated likelihood of existing generated plans. We further present an approach to replan existing trajectories to ensure that new plans follow the same goal state as the original trajectory, which may efficiently bootstrap off previously generated plans. We illustrate how a combination of our proposed additions significantly improves the performance of diffusion planners leading to 38\% gains over past diffusion planning approaches on Maze2D and further enables handling of stochastic and long-horizon robotic control tasks.

We integrate contrastive learning (CL) with adversarial learning to co-optimize the robustness and accuracy of code models. Different from existing works, we show that code obfuscation, a standard code transformation operation, provides novel means to generate complementary ‘views’ of a code that enable us to achieve both robust and accurate code models. To the best of our knowledge, this is the first systematic study to explore and exploit the robustness and accuracy benefits of (multi-view) code obfuscations in code models. Specifically, we first adopt adversarial codes as robustness-promoting views in CL at the self-supervised pre-training phase. This yields improved robustness and transferability for downstream tasks. Next, at the supervised fine-tuning stage, we show that adversarial training with a proper temporally-staggered schedule of adversarial code generation can further improve robustness and accuracy of the pre-trained code model. Built on the above two modules, we develop ClawSAT, a novel self-supervised learning (SSL) framework for code by integrating CL with adversarial views (Claw) with staggered adversarial training (SAT). On evaluating three downstream tasks across Python and Java, we show that ClawSAT consistently yields the best robustness and accuracy (e.g. 11% in robustness and 6% in accuracy on the code summarization task in Python). We additionally demonstrate the effectiveness of adversarial learning in Claw by analyzing the characteristics of the loss landscape and interpretability of the pre-trained models. Codes are available at https://github.com/OPTML-Group/Claw-SAT.

Transformers have proven superior performance for a wide variety of tasks since they were introduced. In recent years, they have drawn attention from the vision community in tasks such as image classification and object detection. Despite this wave, an accurate and efficient multiple-object tracking (MOT) method based on transformers is yet to be designed. We argue that the direct application of a transformer architecture with quadratic complexity and insufficient noise-initialized sparse queries - is not optimal for MOT. We propose TransCenter, a transformer-based MOT architecture with dense representations for accurately tracking all the objects while keeping a reasonable runtime. Methodologically, we propose the use of image-related dense detection queries and efficient sparse tracking queries produced by our carefully designed query learning networks (QLN). On one hand, the dense image-related detection queries allow us to infer targets' locations globally and robustly through dense heatmap outputs. On the other hand, the set of sparse tracking queries efficiently interacts with image features in our TransCenter Decoder to associate object positions through time. As a result, TransCenterexhibits remarkable performance improvements and outperforms by a large margin the current state-of-the-art methods in two standard MOT benchmarks with two tracking settings (public/private). TransCenter is also proven efficient and accurate by an extensive ablation study and, comparisons to more naive alternatives and concurrent works. The code is made publicly available at https://github.com/yihongxu/transcenter.

The way an object looks and sounds provide complementary reflections of its physical properties. In many settings cues from vision and audition arrive asynchronously but must be integrated, as when we hear an object dropped on the floor and then must find it. In this paper, we introduce a setting in which to study multi-modal object localization in 3D virtual environments. An object is dropped somewhere in a room. An embodied robot agent, equipped with camera and microphone, must determine what object has been dropped - and where - by combining audio and visual signals with knowledge of the underlying physics. To study this problem, we have generated a large-scale dataset - the Fallen Objects dataset - that includes 8000 instances of 30 physical object categories in 64 rooms. The dataset uses the ThreeDWorld Platform that can simulate physics-based impact sounds and complex physical interactions between objects in a photorealistic setting. As a first step toward addressing this challenge, we develop a set of embodied agent baselines, based on imitation learning, reinforcement learning, and modular planning, and perform an in-depth analysis of the challenge of this new task. This dataset is publicly available11Project page: http://fallen-object.csail.mit.edu.

This paper studies the problem of fixing malfunctional 3D objects. While previous works focus on building passive perception models to learn the functionality from static 3D objects, we argue that functionality is reckoned with respect to the physical interactions between the object and the user. Given a malfunctional object, humans can perform mental simulations to reason about its functionality and figure out how to fix it. Inspired by this, we propose FixIt, a dataset that contains about 5k poorly-designed 3D physical objects paired with choices to fix them. To mimic humans' mental simulation process, we present FixNet, a novel framework that seamlessly incorporates perception and physical dynamics. Specifically, FixNet consists of a perception module to extract the structured representation from the 3D point cloud, a physical dynamics prediction module to simulate the results of interactions on 3D objects, and a functionality prediction module to evaluate the functionality and choose the correct fix. Experimental results show that our framework outperforms baseline models by a large margin, and can generalize well to objects with similar interaction types. Code and dataset are publicly available11http://fixing-malfunctional.csail.mit.edu.

Training a generalizable 3D part segmentation network is quite challenging but of great importance in real-world applications. To tackle this problem, some works design task-specific solutions by translating human understanding of the task to machine's learning process, which faces the risk of missing the optimal strategy since machines do not necessarily understand in the exact human way. Others try to use conventional task-agnostic approaches designed for domain generalization problems with no task prior knowledge considered. To solve the above issues, we propose AutoGPart, a generic method enabling training generalizable 3D part segmentation networks with the task prior considered. AutoGPart builds a supervision space with geometric prior knowledge encoded, and lets the machine to search for the optimal supervisions from the space for a specific segmentation task automatically. Extensive experiments on three generalizable 3D part segmentation tasks are conducted to demonstrate the effectiveness and versatility of AutoGPart. We demonstrate that the performance of segmentation networks using simple backbones can be significantly improved when trained with supervisions searched by our method.

Human can leverage prior experience and learn novel tasks from a handful of demonstrations. In contrast to offline meta-reinforcement learning, which aims to achieve quick adaptation through better algorithm design, we investigate the effect of architecture inductive bias on the few-shot learning capability. We propose a Prompt-based Decision Transformer (Prompt-DT), which leverages the sequential modeling ability of the Transformer architecture and the prompt framework to achieve few-shot adaptation in offline RL. We design the trajectory prompt, which contains segments of the few-shot demonstrations, and encodes task-specific information to guide policy generation. Our experiments in five MuJoCo control benchmarks show that Prompt-DT is a strong few-shot learner without any extra finetuning on unseen target tasks. Prompt-DT outperforms its variants and strong meta offline RL baselines by a large margin with a trajectory prompt containing only a few timesteps. Prompt-DT is also robust to prompt length changes and can generalize to out-of-distribution (OOD) environments. Project page: \href{https://mxu34.github.io/PromptDT/}{https://mxu34.github.io/PromptDT/}.

We introduce a visually-guided task-and-motion planning benchmark, which we call the ThreeDWorld Trans-port Challenge. In this challenge, an embodied agent is spawned randomly in a simulated physical home environment and required to transport a small set of objects scattered around the house with containers. We build this benchmark challenge using the ThreeDWorld simulation: a virtual 3D environment where all objects respond to physics, and a robot agent can be controlled using a fully physics-driven navigation and interaction API. We evaluate several existing agents on this benchmark. Experimental results suggest that: 1) a pure RL model struggles on this challenge; 2) state-of-the-art hierarchical planning-based agents can transport some objects but are still far from solving this task. We anticipate that this benchmark will empower researchers to develop more intelligent physics-aware robot learning algorithms.

Humans integrate multiple sensory modalities (e.g., visual and audio) to build a causal understanding of the physical world. In this work, we propose a novel type of intrinsic motivation for Reinforcement Learning (RL) that encourages the agent to understand the causal effect of its actions through auditory event prediction. First, we allow the agent to collect a small amount of acoustic data and use K-means to discover underlying auditory event clusters. We then train a neural network to predict the auditory events and use the prediction errors as intrinsic rewards to guide RL exploration. We first conduct proof-of-concept experiments using a set of Atari games for an in-depth analysis of our module. We then apply our model to embodied audio-visual exploration using the Habitat simulator and active exploration with a rolling robot using the ThreeDWorld (TDW) simulator. Experimental results demonstrate the advantages of using audio signals over vision-based models as intrinsic rewards to guide RL explorations.

Existing studies for gait recognition are dominated by in-the-lab scenarios. Since people live in real-world senses, gait recognition in the wild is a more practical problem that has recently attracted the attention of the community of multimedia and computer vision. Current methods that obtain state-of-the-art performance on in-the-lab benchmarks achieve much worse accuracy on the recently proposed in-the-wild datasets because these methods can hardly model the varied temporal dynamics of gait sequences in unconstrained scenes. Therefore, this paper presents a novel multi-hop temporal switch method to achieve effective temporal modeling of gait patterns in real-world scenes. Concretely, we design a novel gait recognition network, named Multi-hop Temporal Switch Network (MTSGait), to learn spatial features and multi-scale temporal features simultaneously. Different from existing methods that use 3D convolutions for temporal modeling, our MTSGait models the temporal dynamics of gait sequences by 2D convolutions. By this means, it achieves high efficiency with fewer model parameters and reduces the difficulty in optimization compared with 3D convolution-based models. Based on the specific design of the 2D convolution kernels, our method can eliminate the misalignment of features among adjacent frames. In addition, a new sampling strategy, i.e., non-cyclic continuous sampling, is proposed to make the model learn more robust temporal features. Finally, the proposed method achieves superior performance on two public gait in-the-wild datasets, i.e., GREW and Gait3D, compared with state-of-the-art methods.

On-device training enables the model to adapt to new data collected from the sensors by fine-tuning a pre-trained model. Users can benefit from customized AI models without having to transfer the data to the cloud, protecting the privacy. However, the training memory consumption is prohibitive for IoT devices that have tiny memory resources. We propose an algorithm-system co-design framework to make on-device training possible with only 256KB of memory. On-device training faces two unique challenges: (1) the quantized graphs of neural networks are hard to optimize due to low bit-precision and the lack of normalization; (2) the limited hardware resource (memory and computation) does not allow full backpropagation. To cope with the optimization difficulty, we propose Quantization- Aware Scaling to calibrate the gradient scales and stabilize 8-bit quantized training. To reduce the memory footprint, we propose Sparse Update to skip the gradient computation of less important layers and sub-tensors. The algorithm innovation is implemented by a lightweight training system, Tiny Training Engine, which prunes the backward computation graph to support sparse updates and offload the runtime auto-differentiation to compile time. Our framework is the first practical solution for on-device transfer learning of visual recognition on tiny IoT devices (e.g., a microcontroller with only 256KB SRAM), using less than 1/1000 of the memory of PyTorch and TensorFlow while matching the accuracy. Our study enables IoT devices not only to perform inference but also to continuously adapt to new data for on-device lifelong learning. A video demo can be found here: https://youtu.be/XaDCO8YtmBw.

Getting robots to navigate to multiple objects autonomously is essential yet difficult in robot applications. One of the key challenges is how to explore environments efficiently with camera sensors only. Existing navigation methods mainly focus on fixed cameras and few attempts have been made to navigate with active cameras. As a result, the agent may take a very long time to perceive the environment due to limited camera scope. In contrast, humans typically gain a larger field of view by looking around for a better perception of the environment. How to make robots perceive the environment as efficiently as humans is a fundamental problem in robotics. In this paper, we consider navigating to multiple objects more efficiently with active cameras. Specifically, we cast moving camera to a Markov Decision Process and reformulate the active camera problem as a reinforcement learning problem. However, we have to address two new challenges: 1) how to learn a good camera policy in complex environments and 2) how to coordinate it with the navigation policy. To address these, we carefully design a reward function to encourage the agent to explore more areas by moving camera actively. Moreover, we exploit human experience to infer a rule-based camera action to guide the learning process. Last, to better coordinate two kinds of policies, the camera policy takes navigation actions into account when making camera moving decisions. Experimental results show our camera policy consistently improves the performance of multi-object navigation over four baselines on two datasets.

We address a practical yet challenging problem of training robot agents to navigate in an environment following a path described by some language instructions. The instructions often contain descriptions of objects in the environment. To achieve accurate and efficient navigation, it is critical to build a map that accurately represents both spatial location and the semantic information of the environment objects. However, enabling a robot to build a map that well represents the environment is extremely challenging as the environment often involves diverse objects with various attributes. In this paper, we propose a multi-granularity map, which contains both object fine-grained details (\eg, color, texture) and semantic classes, to represent objects more comprehensively. Moreover, we propose a weakly-supervised auxiliary task, which requires the agent to localize instruction-relevant objects on the map. Through this task, the agent not only learns to localize the instruction-relevant objects for navigation but also is encouraged to learn a better map representation that reveals object information. We then feed the learned map and instruction to a waypoint predictor to determine the next navigation goal. Experimental results show our method outperforms the state-of-the-art by 4.0% and 4.6% w.r.t. success rate both in seen and unseen environments, respectively on VLN-CE dataset. The code is available at https://github.com/PeihaoChen/WS-MGMap.

Graph Neural Networks (GNNs) have become a prevailing tool for learning physical dynamics. However, they still encounter several challenges: 1) Physical laws abide by symmetry, which is a vital inductive bias accounting for model generalization and should be incorporated into the model design. Existing simulators either consider insufficient symmetry, or enforce excessive equivariance in practice when symmetry is partially broken by gravity. 2) Objects in the physical world possess diverse shapes, sizes, and properties, which should be appropriately processed by the model. To tackle these difficulties, we propose a novel backbone, called Subequivariant Graph Neural Network, which 1) relaxes equivariance to subequivariance by considering external fields like gravity, where the universal approximation ability holds theoretically; 2) introduces a new subequivariant object-aware message passing for learning physical interactions between multiple objects of various shapes in particle-based representation; 3) operates in a hierarchical fashion, allowing for modeling long-range and complex interactions. Our model achieves on average over 3% enhancement in contact prediction accuracy across 8 scenarios on Physion and 2$\times$ lower rollout MSE on RigidFall compared with state-of-the-art GNN simulators, while exhibiting strong generalization and data efficiency.

In this paper, we address the challenging problem of 3D concept grounding (i.e., segmenting and learning visual concepts) by looking at RGBD images and reasoning about paired questions and answers. Existing visual reasoning approaches typically utilize supervised methods to extract 2D segmentation masks on which concepts are grounded. In contrast, humans are capable of grounding concepts on the underlying 3D representation of images. However, traditionally inferred 3D representations (e.g., point clouds, voxelgrids and meshes) cannot capture continuous 3D features flexibly, thus making it challenging to ground concepts to 3D regions based on the language description of the object being referred to. To address both issues, we propose to leverage the continuous, differentiable nature of neural fields to segment and learn concepts. Specifically, each 3D coordinate in a scene is represented as a high dimensional descriptor. Concept grounding can then be performed by computing the similarity between the descriptor vector of a 3D coordinate and the vector embedding of a language concept, which enables segmentations and concept learning to be jointly learned on neural fields in a differentiable fashion. As a result, both 3D semantic and instance segmentations can emerge directly from question answering supervision using a set of defined neural operators on top of neural fields (e.g., filtering and counting). Experimental results show that our proposed framework outperforms unsupervised / language-mediated segmentation models on semantic and instance segmentation tasks, as well as outperforms existing models on the challenging 3D aware visual reasoning tasks. Furthermore, our framework can generalize well to unseen shape categories and real scans.

Our environment is filled with rich and dynamic acoustic information. When we walk into a cathedral, the reverberations as much as appearance inform us of the sanctuary's wide open space. Similarly, as an object moves around us, we expect the sound emitted to also exhibit this movement. While recent advances in learned implicit functions have led to increasingly higher quality representations of the visual world, there have not been commensurate advances in learning spatial auditory representations. To address this gap, we introduce Neural Acoustic Fields (NAFs), an implicit representation that captures how sounds propagate in a physical scene. By modeling acoustic propagation in a scene as a linear time-invariant system, NAFs learn to continuously map all emitter and listener location pairs to a neural impulse response function that can then be applied to arbitrary sounds. We demonstrate NAFs on both synthetic and real data, and show that the continuous nature of NAFs enables us to render spatial acoustics for a listener at arbitrary locations. We further show that the representation learned by NAFs can help improve visual learning with sparse views. Finally we show that a representation informative of scene structure emerges during the learning of NAFs.

Detecting 3D keypoints from point clouds is important for shape reconstruction, while this work investigates the dual question: can shape reconstruction benefit 3D keypoint detection? Existing methods either seek salient features according to statistics of different orders or learn to predict keypoints that are invariant to transformation. Nevertheless, the idea of incorporating shape reconstruction into 3D keypoint detection is under-explored. We argue that this is restricted by former problem formulations. To this end, a novel unsupervised paradigm named SNAKE is proposed, which is short for shape-aware neural 3D keypoint field. Similar to recent coordinate-based radiance or distance field, our network takes 3D coordinates as inputs and predicts implicit shape indicators and keypoint saliency simultaneously, thus naturally entangling 3D keypoint detection and shape reconstruction. We achieve superior performance on various public benchmarks, including standalone object datasets ModelNet40, KeypointNet, SMPL meshes and scene-level datasets 3DMatch and Redwood. Intrinsic shape awareness brings several advantages as follows. (1) SNAKE generates 3D keypoints consistent with human semantic annotation, even without such supervision. (2) SNAKE outperforms counterparts in terms of repeatability, especially when the input point clouds are down-sampled. (3) the generated keypoints allow accurate geometric registration, notably in a zero-shot setting. Codes and models are available at https://github.com/zhongcl-thu/SNAKE.

The explosive growth in video streaming requires video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. The key idea of TSM is to shift part of the channels along the temporal dimension, thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. TSM offers several unique advantages. First, TSM has high performance; it ranks the first on the Something-Something leaderboard upon submission. Second, TSM has high efficiency; it achieves a high frame rate of 74fps and 29fps for online video recognition on Jetson Nano and Galaxy Note8. Third, TSM has higher scalability compared to 3D networks, enabling large-scale Kinetics training on 1,536 GPUs in 15 minutes. Lastly, TSM enables action concepts learning, which 2D networks cannot model; we visualize the category attention map and find that spatial-temporal action detector emerges during the training of classification tasks. The code is publicly available at https://github.com/mit-han-lab/temporal-shift-module.

Recently, substantial research effort has focused on how to apply CNNs or RNNs to better capture temporal patterns in videos, so as to improve the accuracy of video classification. In this paper, we investigate the potential of a purely attention based local feature integration. Accounting for the characteristics of such features in video classification, we first propose Basic Attention Clusters (BAC), which concatenates the output of multiple attention units applied in parallel, and introduce a shifting operation to capture more diverse signals. Experiments show that BAC can achieve excellent results on multiple datasets. However, BAC treats all feature channels as an indivisible whole, which is suboptimal for achieving a finer-grained local feature integration over the channel dimension. Additionally, it treats the entire local feature sequence as an unordered set, thus ignoring the sequential relationships. To improve over BAC, we further propose the channel pyramid attention schema by splitting features into sub-features at multiple scales for coarse-to-fine sub-feature interaction modeling, and propose the temporal pyramid attention schema by dividing the feature sequences into ordered sub-sequences of multiple lengths to account for the sequential order. Our final model pyramid×pyramid attention clusters (PPAC) combines both channel pyramid attention and temporal pyramid attention to focus on the most important sub-features, while also preserving the temporal information of the video. We demonstrate the effectiveness of PPAC on seven real-world video classification datasets. Our model achieves competitive results across all of these, showing that our proposed framework can consistently outperform the existing local feature integration methods across a range of different scenarios.

Temporal action localization, which requires a machine to recognize the location as well as the category of action instances in videos, has long been researched in computer vision. The main challenge of temporal action localization lies in that videos are usually long and untrimmed with diverse action contents involved. Existing state-of-the-art action localization methods divide each video into multiple action units (i.e., proposals in two-stage methods and segments in one-stage methods) and then perform action recognition/regression on each of them individually, without explicitly exploiting their relations during learning. In this paper, we claim that the relations between action units play an important role in action localization, and a more powerful action detector should not only capture the local content of each action unit but also allow a wider field of view on the context related to it. To this end, we propose a general graph convolutional module (GCM) that can be easily plugged into existing action localization methods, including two-stage and one-stage paradigms. To be specific, we first construct a graph, where each action unit is represented as a node and their relations between two action units as an edge. Here, we use two types of relations, one for capturing the temporal connections between different action units, and the other one for characterizing their semantic relationship. Particularly for the temporal connections in two-stage methods, we further explore two different kinds of edges, one connecting the overlapping action units and the other one connecting surrounding but disjointed units. Upon the graph we built, we then apply graph convolutional networks (GCNs) to model the relations among different action units, which is able to learn more informative representations to enhance action localization. Experimental results show that our GCM consistently improves the performance of existing action localization methods, including two-stage methods (e.g., CBR [15] and R-C3D [47]) and one-stage methods (e.g., D-SSAD [22]), verifying the generality and effectiveness of our GCM. Moreover, with the aid of GCM, our approach significantly outperforms the state-of-the-art on THUMOS14 (50.9 percent versus 42.8 percent). Augmentation experiments on ActivityNet also verify the efficacy of modeling the relationships between action units. The source code and the pre-trained models are available at https://github.com/Alvin-Zeng/GCM.

We study unsupervised video representation learning that seeks to learn both motion and appearance features from unlabeled video only, which can be reused for downstream tasks such as action recognition. This task, however, is extremely challenging due to 1) the highly complex spatial-temporal information in videos and 2) the lack of labeled data for training. Unlike representation learning for static images, it is difficult to construct a suitable self-supervised task to effectively model both motion and appearance features. More recently, several attempts have been made to learn video representation through video playback speed prediction. However, it is non-trivial to obtain precise speed labels for the videos. More critically, the learned models may tend to focus on motion patterns and thus may not learn appearance features well. In this paper, we observe that the relative playback speed is more consistent with motion patterns and thus provides more effective and stable supervision for representation learning. Therefore, we propose a new way to perceive the playback speed and exploit the relative speed between two video clips as labels. In this way, we are able to effectively perceive speed and learn better motion features. Moreover, to ensure the learning of appearance features, we further propose an appearance-focused task, where we enforce the model to perceive the appearance difference between two video clips. We show that jointly optimizing the two tasks consistently improves the performance on two downstream tasks (namely, action recognition and video retrieval) w.r.t the increasing pre-training epochs. Remarkably, for action recognition on the UCF101 dataset, we achieve 93.7% accuracy without the use of labeled data for pre-training, which outperforms the ImageNet supervised pre-trained model. Our code, pre-trained models, and supplementary materials can be found at https://github.com/PeihaoChen/RSPNet.

Conventionally, spatiotemporal modeling network and its complexity are the two most concentrated research topics in video action recognition. Existing state-of-the-art methods have achieved excellent accuracy regardless of the complexity meanwhile efficient spatiotemporal modeling solutions are slightly inferior in performance. In this paper, we attempt to acquire both efficiency and effectiveness simultaneously. First of all, besides traditionally treating H x W x T video frames as space-time signal (viewing from the Height-Width spatial plane), we propose to also model video from the other two Height-Time and Width-Time planes, to capture the dynamics of video thoroughly. Secondly, our model is designed based on 2D CNN backbones and model complexity is well kept in mind by design. Specifically, we introduce a novel multi-view fusion (MVF) module to exploit video dynamics using separable convolution for efficiency. It is a plug-and-play module and can be inserted into off-the-shelf 2D CNNs to form a simple yet effective model called MVFNet. Moreover, MVFNet can be thought of as a generalized video modeling framework and it can specialize to be existing methods such as C2D, SlowOnly, and TSM under different settings. Extensive experiments are conducted on popular benchmarks (i.e., Something-Something V1 & V2, Kinetics, UCF-101, and HMDB-51) to show its superiority. The proposed MVFNet can achieve state-of-the-art performance with 2D CNN's complexity.

Humans can abstract prior knowledge from very little data and use it to boost skill learning. In this paper, we propose routine-augmented policy learning (RAPL), which discovers routines composed of primitive actions from a single demonstration and uses discovered routines to augment policy learning. To discover routines from the demonstration, we first abstract routine candidates by identifying grammar over the demonstrated action trajectory. Then, the best routines measured by length and frequency are selected to form a routine library. We propose to learn policy simultaneously at primitive-level and routine-level with discovered routines, leveraging the temporal structure of routines. Our approach enables imitating expert behavior at multiple temporal scales for imitation learning and promotes reinforcement learning exploration. Extensive experiments on Atari games demonstrate that RAPL improves the state-of-the-art imitation learning method SQIL and reinforcement learning method A2C. Further, we show that discovered routines can generalize to unseen levels and difficulties on the CoinRun benchmark.

The problem of grounding VQA tasks has seen an increased attention in the research community recently, with most attempts usually focusing on solving this task by using pretrained object detectors. However, pre-trained object detectors require bounding box annotations for detecting relevant objects in the vocabulary, which may not always be feasible for real-life large-scale applications. In this paper, we focus on a more relaxed setting: the grounding of relevant visual entities in a weakly supervised manner by training on the VQA task alone. To address this problem, we propose a visual capsule module with a query-based selection mechanism of capsule features, that allows the model to focus on relevant regions based on the textual cues about visual information in the question. We show that integrating the proposed capsule module in existing VQA systems significantly improves their performance on the weakly supervised grounding task. Overall, we demonstrate the effectiveness of our approach on two state-of-the-art VQA systems, stacked NMN and MAC, on the CLEVR-Answers benchmark, our new evaluation set based on CLEVR scenes with groundtruth bounding boxes for objects that are relevant for the correct answer, as well as on GQA, a real world VQA dataset with compositional questions. We show that the systems with the proposed capsule module consistently outperform the respective baseline systems in terms of answer grounding, while achieving comparable performance on VQA task.1

Self-supervised representation learning has achieved remarkable success in recent years. By subverting the need for supervised labels, such approaches are able to utilize the numerous unlabeled images that exist on the Internet and in photographic datasets. Yet to build truly intelligent agents, we must construct representation learning algorithms that can learn not only from datasets but also learn from environments. An agent in a natural environment will not typically be fed curated data. Instead, it must explore its environment to acquire the data it will learn from. We propose a framework, curious representation learning (CRL), which jointly learns a reinforcement learning policy and a visual representation model. The policy is trained to maximize the error of the representation learner, and in doing so is incentivized to explore its environment. At the same time, the learned representation becomes stronger and stronger as the policy feeds it ever harder data to learn from. Our learned representations enable promising transfer to downstream navigation tasks, performing better than or comparably to ImageNet pretraining without using any supervision at all. In addition, despite being trained in simulation, our learned representations can obtain interpretable results on real images. Code is available at https://yilundu.github.io/crl/.

It has been a challenge to learning skills for an agent from long-horizon unannotated demonstrations. Existing approaches like Hierarchical Imitation Learning(HIL) are prone to compounding errors or suboptimal solutions. In this paper, we propose Option-GAIL, a novel method to learn skills at long horizon. The key idea of Option-GAIL is modeling the task hierarchy by options and train the policy via generative adversarial optimization. In particular, we propose an Expectation-Maximization(EM)-style algorithm: an E-step that samples the options of expert conditioned on the current learned policy, and an M-step that updates the low- and high-level policies of agent simultaneously to minimize the newly proposed option-occupancy measurement between the expert and the agent. We theoretically prove the convergence of the proposed algorithm. Experiments show that Option-GAIL outperforms other counterparts consistently across a variety of tasks.

Prosody plays an important role in characterizing the style of a speaker or an emotion, but most non-parallel voice or emotion style transfer algorithms do not convert any prosody information. Two major components of prosody are pitch and rhythm. Disentangling the prosody information, particularly the rhythm component, from the speech is challenging because it involves breaking the synchrony between the input speech and the disentangled speech representation. As a result, most existing prosody style transfer algorithms would need to rely on some form of text transcriptions to identify the content information, which confines their application to high-resource languages only. Recently, SpeechSplit has made sizeable progress towards unsupervised prosody style transfer, but it is unable to extract high-level global prosody style in an unsupervised manner. In this paper, we propose AutoPST, which can disentangle global prosody style from speech without relying on any text transcriptions. AutoPST is an Autoencoder-based Prosody Style Transfer framework with a thorough rhythm removal module guided by the self-expressive representation learning. Experiments on different style transfer tasks show that AutoPST can effectively convert prosody that correctly reflects the styles of the target domains.

For machine agents to successfully interact with humans in real-world settings, they will need to develop an understanding of human mental life. Intuitive psychology, the ability to reason about hidden mental variables that drive observable actions, comes naturally to people: even pre-verbal infants can tell agents from objects, expecting agents to act efficiently to achieve goals given constraints. Despite recent interest in machine agents that reason about other agents, it is not clear if such agents learn or hold the core psychology principles that drive human reasoning. Inspired by cognitive development studies on intuitive psychology, we present a benchmark consisting of a large dataset of procedurally generated 3D animations, AGENT (Action, Goal, Efficiency, coNstraint, uTility), structured around four scenarios (goal preferences, action efficiency, unobserved constraints, and cost-reward trade-offs) that probe key concepts of core intuitive psychology. We validate AGENT with human-ratings, propose an evaluation protocol emphasizing generalization, and compare two strong baselines built on Bayesian inverse planning and a Theory of Mind neural network. Our results suggest that to pass the designed tests of core intuitive psychology at human levels, a model must acquire or have built-in representations of how agents plan, combining utility computations and core knowledge of objects and physics.

We present Temporal and Object Quantification Networks (TOQ-Nets), a new class of neuro-symbolic networks with a structural bias that enables them to learn to recognize complex relational-temporal events. This is done by including reasoning layers that implement finite-domain quantification over objects and time. The structure allows them to generalize directly to input instances with varying numbers of objects in temporal sequences of varying lengths. We evaluate TOQ-Nets on input domains that require recognizing event-types in terms of complex temporal relational patterns. We demonstrate that TOQ-Nets can generalize from small amounts of data to scenarios containing more objects than were present during training and to temporal warpings of input sequences.

Humans have mental models that allow them to plan, experiment, and reason in the physical world. How should an intelligent agent go about learning such models? In this paper, we will study if models of the world learned in an open-ended physics environment, without any specific tasks, can be reused for downstream physics reasoning tasks. To this end, we build a benchmark Open-ended Physics Environment (OPEn) and also design several tasks to test learning representations in this environment explicitly. This setting reflects the conditions in which real agents (i.e. rolling robots) find themselves, where they may be placed in a new kind of environment and must adapt without any teacher to tell them how this environment works. This setting is challenging because it requires solving an exploration problem in addition to a model building and representation learning problem. We test several existing RL-based exploration methods on this benchmark and find that an agent using unsupervised contrastive learning for representation learning, and impact-driven learning for exploration, achieved the best results. However, all models still fall short in sample efficiency when transferring to the downstream tasks. We expect that OPEn will encourage the development of novel rolling robot agents that can build reusable mental models of the world that facilitate many tasks.

Compositional action recognition is a novel challenge in the computer vision community and focuses on revealing the different combinations of verbs and nouns instead of treating subject-object interactions in videos as individual instances only. Existing methods tackle this challenging task by simply ignoring appearance information or fusing object appearances with dynamic instance tracklets. However, those strategies usually do not perform well for unseen action instances. For that, in this work we propose a novel learning framework called Counterfactual Debiasing Network (CDN) to improve the model generalization ability by removing the interference introduced by visual appearances of objects/subjects. It explicitly learns the appearance information in action representations and later removes the effect of such information in a causal inference manner. Specifically, we use tracklets and video content to model the factual inference by considering both appearance information and structure information. In contrast, only video content with appearance information is leveraged in the counterfactual inference. With the two inferences, we conduct a causal graph which captures and removes the bias introduced by the appearance information by subtracting the result of the counterfactual inference from that of the factual inference. By doing that, our proposed CDN method can better recognize unseen action instances by debiasing the effect of appearances. Extensive experiments on the Something-Else dataset clearly show the effectiveness of our proposed CDN over existing state-of-the-art methods.

Skeleton-based action recognition has been widely investigated considering their strong adaptability to dynamic circumstances and complicated backgrounds. To recognize different actions from skeleton sequences, it is essential and crucial to model the posture of the human represented by the skeleton and its changes in the temporal dimension. However, most of the existing works treat skeleton sequences in the temporal and spatial dimension in the same way, ignoring the difference between the temporal and spatial dimension in skeleton data which is not an optimal way to model skeleton sequences. The posture represented by the skeleton in each frame is proposed to be modeled individually. Meanwhile, capturing the movement of the entire skeleton in the temporal dimension is needed. So, we designed Spatial Transformer Block and Directional Temporal Transformer Block for modeling skeleton sequences in spatial and temporal dimensions respectively. Due to occlusion/sensor/raw video, etc., there are noises on both temporal and spatial dimensions in the extracted skeleton data reducing the recognition capabilities of models. To adapt to this imperfect information condition, we propose a multi-task self-supervised learning method by providing confusing samples in different situations to improve the robustness of our model. Combining the above design, we propose our Spatial-Temporal Specialized Transformer~(STST) and conduct experiments with our model on the SHREC, NTU-RGB+D, and Kinetics-Skeleton. Extensive experimental results demonstrate the improved performances and analysis of the proposed method.

In this work, we propose a unified framework, called Visual Reasoning with Differ-entiable Physics (VRDP), that can jointly learn visual concepts and infer physics models of objects and their interactions from videos and language. This is achieved by seamlessly integrating three components: a visual perception module, a concept learner, and a differentiable physics engine. The visual perception module parses each video frame into object-centric trajectories and represents them as latent scene representations. The concept learner grounds visual concepts (e.g., color, shape, and material) from these object-centric representations based on the language, thus providing prior knowledge for the physics engine. The differentiable physics model, implemented as an impulse-based differentiable rigid-body simulator, performs differentiable physical simulation based on the grounded concepts to infer physical properties, such as mass, restitution, and velocity, by fitting the simulated trajectories into the video observations. Consequently, these learned concepts and physical models can explain what we have seen and imagine what is about to happen in future and counterfactual scenarios. Integrating differentiable physics into the dynamic reasoning framework offers several appealing benefits. More accurate dynamics prediction in learned physics models enables state-of-the-art performance on both synthetic and real-world benchmarks while still maintaining high transparency and interpretability; most notably, VRDP improves the accuracy of predictive and counterfactual questions by 4.5% and 11.5% compared to its best counterpart. VRDP is also highly data-efficient: physical parameters can be optimized from very few videos, and even a single video can be sufficient. Finally, with all physical parameters inferred, VRDP can quickly learn new concepts from a few examples.

Contrastive learning (CL) can learn generalizable feature representations and achieve state-of-the-art performance of downstream tasks by finetuning a linear classifier on top of it. However, as adversarial robustness becomes vital in image classification, it remains unclear whether or not CL is able to preserve robustness to downstream tasks. The main challenge is that in the self-supervised pretraining + supervised finetuning paradigm, adversarial robustness is easily forgotten due to a learning task mismatch from pretraining to finetuning. We call such challenge 'cross-task robustness transferability'. To address the above problem, in this paper we revisit and advance CL principles through the lens of robustness enhancement. We show that (1) the design of contrastive views matters: High-frequency components of images are beneficial to improving model robustness; (2) Augmenting CL with pseudo-supervision stimulus (e.g., resorting to feature clustering) helps preserve robustness without forgetting. Equipped with our new designs, we propose AdvCL, a novel adversarial contrastive pretraining framework. We show that AdvCL is able to enhance cross-task robustness transferability without loss of model accuracy and finetuning efficiency. With a thorough experimental study, we demonstrate that AdvCL outperforms the state-of-the-art self-supervised robust learning methods across multiple datasets (CIFAR-10, CIFAR-100, and STL-10) and finetuning schemes (linear evaluation and full model finetuning).

A critical aspect of human visual perception is the ability to parse visual scenes into individual objects and further into object parts, forming part-whole hierarchies. Such composite structures could induce a rich set of semantic concepts and relations, thus playing an important role in the interpretation and organization of visual signals as well as for the generalization of visual perception and reasoning. However, existing visual reasoning benchmarks mostly focus on objects rather than parts. Visual reasoning based on the full part-whole hierarchy is much more challenging than object-centric reasoning due to finer-grained concepts, richer geometry relations, and more complex physics. Therefore, to better serve for part-based conceptual, relational and physical reasoning, we introduce a new large-scale diagnostic visual reasoning dataset named PTR. PTR contains around 80k RGBD synthetic images with ground truth object and part level annotations regarding semantic instance segmentation, color attributes, spatial and geometric relationships, and certain physical properties such as stability. These images are paired with 800k machine-generated questions covering various types of reasoning types, making them a good testbed for visual reasoning models. We examine several state-of-the-art visual reasoning models on this dataset and observe that they still make many surprising mistakes in situations where humans can easily infer the correct answer. We believe this dataset will open up new opportunities for part-based reasoning. PTR dataset and baseline models are publicly available.

Tiny deep learning on microcontroller units (MCUs) is challenging due to the limited memory size. We find that the memory bottleneck is due to the imbalanced memory distribution in convolutional neural network (CNN) designs: the first several blocks have an order of magnitude larger memory usage than the rest of the network. To alleviate this issue, we propose a generic patch-by-patch inference scheduling, which operates only on a small spatial region of the feature map and significantly cuts down the peak memory. However, naive implementation brings overlapping patches and computation overhead. We further propose receptive field redistribution to shift the receptive field and FLOPs to the later stage and reduce the computation overhead. Manually redistributing the receptive field is difficult. We automate the process with neural architecture search to jointly optimize the neural architecture and inference scheduling, leading to MCUNetV2. Patch-based inference effectively reduces the peak memory usage of existing networks by4-8×. Co-designed with neural networks, MCUNetV2 sets a record ImageNetaccuracy on MCU (71.8%) and achieves >90% accuracy on the visual wake words dataset under only 32kB SRAM. MCUNetV2 also unblocks object detection on tiny devices, achieving 16.9% higher mAP on Pascal VOC compared to the state-of-the-art result. Our study largely addressed the memory bottleneck in tinyML and paved the way for various vision applications beyond image classification.

Recent years have witnessed the success of machine learning and especially deep learning in many research areas such as Vision and Language Processing, Information Retrieval and Recommender Systems, Social Networks and Conversational Agents. Though various learning approaches have demonstrated satisfying performance in perceptual tasks such as associative learning and matching by extracting useful similarity patterns from data, the area still sees a large amount of research needed to advance the ability of reasoning towards cognitive intelligence in the coming years. This includes but is not limited to neural logical reasoning, neural-symbolic reasoning, causal reasoning, knowledge reasoning and commonsense reasoning. The workshop focuses on the research of machine reasoning techniques and their application in various intelligent tasks. It will gather researchers as well as practitioners in the field for discussions, idea communications, and research promotions. It will also generate insightful debates about the recent progress in machine intelligence to a broader community, including but not limited to CV, IR, NLP, ML, DM, AI and beyond.

Deep neural networks have achieved remarkable success for video-based action recognition. However, most of existing approaches cannot be deployed in practice due to the high computational cost. To address this challenge, we propose a new real-time convolutional architecture, called Temporal Convolutional 3D Network (T-C3D), for action representation. T-C3D learns video action representations in a hierarchical multi-granularity manner while obtaining a high process speed. Specifically, we propose a residual 3D Convolutional Neural Network (CNN) to capture complementary information on the appearance of a single frame and the motion between consecutive frames. Based on this CNN, we develop a new temporal encoding method to explore the temporal dynamics of the whole video. Furthermore, we integrate deep compression techniques with T-C3D to further accelerate the deployment of models via reducing the size of the model. By these means, heavy calculations can be avoided when doing the inference, which enables the method to deal with videos beyond real-time speed while keeping promising performance. We validate our approach by studying its action representation performance on four benchmarks over three different tasks. Our method achieves clear improvements on UCF101 action recognition benchmark against the state-of-the-art real-time methods by 5.4% in terms of accuracy and 2 times faster in terms of inference speed with a less than 5MB storage model. The source code and the pre-trained models are publicly available at https://github.com/tc3d.

We addressed the challenging task of video question answering, which requires machines to answer questions about videos in a natural language form. Previous state-of-the-art methods attempt to apply spatio-temporal attention mechanism on video frame features without explicitly modeling the location and relations among object interaction occurred in videos. However, the relations between object interaction and their location information are very critical for both action recognition and question reasoning. In this work, we propose to represent the contents in the video as a location-aware graph by incorporating the location information of an object into the graph construction. Here, each node is associated with an object represented by its appearance and location features. Based on the constructed graph, we propose to use graph convolution to infer both the category and temporal locations of an action. As the graph is built on objects, our method is able to focus on the foreground action contents for better video question answering. Lastly, we leverage an attention mechanism to combine the output of graph convolution and encoded question features for final answer reasoning. Extensive experiments demonstrate the effectiveness of the proposed methods. Specifically, our method significantly outperforms state-of-the-art methods on TGIF-QA, Youtube2Text-QA and MSVD-QA datasets.

Transformers are ubiquitous in Natural Language Processing (NLP) tasks, but they are difficult to be deployed on hardware due to the intensive computation.To enable low-latency inference on resource-constrained hardware platforms, we propose to design Hardware-Aware Transformers (HAT) with neural architecture search.We first construct a large design space with arbitrary encoder-decoder attention and heterogeneous layers.Then we train a Super-Transformer that covers all candidates in the design space, and efficiently produces many SubTransformers with weight sharing.Finally, we perform an evolutionary search with a hardware latency constraint to find a specialized SubTransformer dedicated to run fast on the target hardware.Extensive experiments on four machine translation tasks demonstrate that HAT can discover efficient models for different hardware (CPU, GPU, IoT device).When running WMT'14 translation task on Raspberry Pi-4, HAT can achieve 3× speedup, 3.7× smaller size over baseline Transformer; 2.7× speedup, 3.6× smaller size over Evolved Transformer with 12,041× less search cost and no performance loss.HAT is open-sourced.Elastic Layer Num in Encoder Elastic Head Num (Self Attention) Elastic Hidden Dim in FFN Encoder Layer 2 Encoder Layer m Elastic Embedding Dim Elastic Head Num (Self Attention) Elastic Hidden Dim in FFN Elastic Embedding Dim Elastic Head Num (En-Decoder Attention) Decoder Layer n Elastic Layer Num in Decoder Arbitrary Encoder-Decoder Attention concat ❶ Train a SuperTransformer by uniformly sampling SubTransformers with weight sharing 12/3/

Recent deep learning approaches have achieved impressive performance on visual sound separation tasks. However, these approaches are mostly built on appearance and optical flow like motion feature representations, which exhibit limited abilities to find the correlations between audio signals and visual points, especially when separating multiple instruments of the same types, such as multiple violins in a scene. To address this, we propose ``Music Gesture," a keypoint-based structured representation to explicitly model the body and finger movements of musicians when they perform music. We first adopt a context-aware graph network to integrate visual semantic context with body dynamics and then apply an audio-visual fusion model to associate body movements with the corresponding audio signals. Experimental results on three music performance datasets show: 1) strong improvements upon benchmark metrics for hetero-musical separation tasks (i.e. different instruments); 2) new ability for effective homo-musical separation for piano, flute, and trumpet duets, which to our best knowledge has never been achieved with alternative methods.

We address the problem of video grounding from natural language queries. The key challenge in this task is that one training video might only contain a few annotated starting/ending frames that can be used as positive examples for model training. Most conventional approaches directly train a binary classifier using such imbalance data, thus achieving inferior results. The key idea of this paper is to use the distances between the frame within the ground truth and the starting (ending) frame as dense supervisions to improve the video grounding accuracy. Specifically, we design a novel dense regression network (DRN) to regress the distances from each frame to the starting (ending) frame of the video segment described by the query. We also propose a simple but effective IoU regression head module to explicitly consider the localization quality of the grounding results (i.e., the IoU between the predicted location and the ground truth). Experimental results show that our approach significantly outperforms state-of-the-arts on three datasets (i.e., Charades-STA, ActivityNet-Captions, and TACoS).

Interactive Fiction (IF) games with real humanwritten natural language texts provide a new natural evaluation for language understanding techniques.In contrast to previous text games with mostly synthetic texts, IF games pose language understanding challenges on the humanwritten textual descriptions of diverse and sophisticated game worlds and language generation challenges on the action command generation from less restricted combinatorial space.We take a novel perspective of IF game solving and re-formulate it as Multi-Passage Reading Comprehension (MPRC) tasks.Our approaches utilize the context-query attention mechanisms and the structured prediction in MPRC to efficiently generate and evaluate action outputs and apply an object-centric historical observation retrieval strategy to mitigate the partial observability of the textual observations.Extensive experiments on the recent IF benchmark (Jericho) demonstrate clear advantages of our approaches achieving high winning rates and low data requirements compared to all previous approaches. 1

A crucial ability of mobile intelligent agents is to integrate the evidence from multiple sensory inputs in an environment and to make a sequence of actions to reach their goals. In this paper, we attempt to approach the problem of Audio-Visual Embodied Navigation, the task of planning the shortest path from a random starting location in a scene to the sound source in an indoor environment, given only raw egocentric visual and audio sensory data. To accomplish this task, the agent is required to learn from various modalities, i.e., relating the audio signal to the visual environment. Here we describe an approach to audio-visual embodied navigation that takes advantage of both visual and audio pieces of evidence. Our solution is based on three key ideas: a visual perception mapper module that constructs its spatial memory of the environment, a sound perception module that infers the relative location of the sound source from the agent, and a dynamic path planner that plans a sequence of actions based on the audio-visual observations and the spatial memory of the environment to navigate toward the goal. Experimental results on a newly collected Visual-Audio-Room dataset using the simulated multi-modal environment demonstrate the effectiveness of our approach over several competitive baselines.

Existing action localization approaches adopt shallow temporal convolutional networks (i.e., TCN) on 1D feature map extracted from video frames. In this paper, we empirically find that stacking more conventional temporal convolution layers actually deteriorates action classification performance, possibly ascribing to that all channels of 1D feature map, which generally are highly abstract and can be regarded as latent concepts, are excessively recombined in temporal convolution. To address this issue, we introduce a novel concept-wise temporal convolutional network (C-TCN) as an alternative to TCN for training deeper action localization networks. To address this issue, we introduce a novel concept-wise temporal convolution (CTC) layer as an alternative to conventional temporal convolution layer for training deeper action localization networks. Instead of recombining latent concepts, CTC layer deploys a number of temporal filters to each concept separately with shared filter parameters across concepts. Thus can capture common temporal patterns of different concepts and significantly enrich representation ability. Via stacking CTC layers, we proposed a deep concept-wise temporal convolutional network (C-TCN), which boosts the state-of-the-art action localization performance on THUMOS'14 from 42.8 to 52.1 in terms of mAP(%), achieving a relative improvement of 21.7%. Favorable result is also obtained on ActivityNet.

The First International Workshop on Human-Centric MultimediaAnalysis is concentrated on the tasks of human-centric analysis with multimedia and multimodal information. It is one of the fundamental and challenging problems of multimedia understanding. The human-centric multimedia analysis involves multiple tasks such as face detection and recognition, human body pattern analysis, person re-identification, human action detection, person tracking,human-object interaction, and so on. Today, multiple multimedia sensing technologies and large-scale computing infrastructures are producing at a rapid velocity a wide variety of big multi-modality data for human-centric analysis, which provides rich knowledge to help tackle these challenges. Researchers have strived to push the limits of human-centric multimedia analysis in a wide variety of applications, such as intelligent surveillance, retailing, fashion design, and services. Therefore, this workshop aims to provide a platform to bridge the gap between the communities of human analysis and multimedia.

We address the challenging task of event localization, which requires the machine to localize an event and recognize its category in unconstrained videos. Most existing methods leverage only the visual information of a video while neglecting its audio information, which, however, can be very helpful and important for event localization. For example, humans often recognize an event by reasoning with the visual and audio content simultaneously. Moreover, the audio information can guide the model to pay more attention on the informative regions of visual scenes, which can help to reduce the interference brought by the background. Motivated by these, in this paper, we propose a relation-aware network to leverage both audio and visual information for accurate event localization. Specifically, to reduce the interference brought by the background, we propose an audio-guided spatial-channel attention module to guide the model to focus on event-relevant visual regions. Besides, we propose to build connections between visual and audio modalities with a relation-aware module. In particular, we learn the representations of video and/or audio segments by aggregating information from the other modality according to the cross-modal relations. Last, relying on the relation-aware representations, we conduct event localization by predicting the event relevant score and classification score. Extensive experimental results demonstrate that our method significantly outperforms the state-of-the-arts in both supervised and weakly-supervised AVE settings.

Efficient on-device learning requires a small memory footprint at training time to fit the tight memory constraint. Existing work solves this problem by reducing the number of trainable parameters. However, this doesn't directly translate to memory saving since the major bottleneck is the activations, not parameters. In this work, we present Tiny-Transfer-Learning (TinyTL) for memory-efficient on-device learning. TinyTL freezes the weights while only learns the memory-efficient bias modules, thus no need to store the intermediate activations. To maintain the adaptation capacity, we introduce a new memory-efficient bias module, the lite residual module, to refine the feature extractor by learning small residual feature maps adding only 3.8% memory overhead. Extensive experiments show that TinyTL significantly saves the memory (up to 6.5x) with little accuracy loss compared to fine-tuning the full network. Compared to fine-tuning the last layer, TinyTL provides significant accuracy improvements (up to 33.8%) with little memory overhead. Furthermore, combined with feature extractor adaptation, TinyTL provides 7.5-12.9x memory saving without sacrificing accuracy compared to fine-tuning the full Inception-V3. Code is released at https://github.com/mit-han-lab/tinyML/tree/master/tinyTL.

Machine learning on tiny IoT devices based on microcontroller units (MCU) is appealing but challenging: the memory of microcontrollers is 2-3 orders of magnitude smaller even than mobile phones. We propose MCUNet, a framework that jointly designs the efficient neural architecture (TinyNAS) and the lightweight inference engine (TinyEngine), enabling ImageNet-scale inference on microcontrollers. TinyNAS adopts a two-stage neural architecture search approach that first optimizes the search space to fit the resource constraints, then specializes the network architecture in the optimized search space. TinyNAS can automatically handle diverse constraints (i.e. device, latency, energy, memory) under low search costs. TinyNAS is co-designed with TinyEngine, a memory-efficient inference library to expand the search space and fit a larger model. TinyEngine adapts the memory scheduling according to the overall network topology rather than layer-wise optimization, reducing the memory usage by 3.4×, and accelerating the inference by 1.7-3.3× compared to TF-Lite Micro [3] and CMSIS-NN [28]. MCUNet is the first to achieves >70% ImageNet top1 accuracy on an off-the-shelf commercial microcontroller, using 3.5× less SRAM and 5.7× less Flash compared to quantized MobileNetV2 and ResNet-18. On visual&audio wake words tasks, MCUNet achieves state-of-the-art accuracy and runs 2.4-3.4× faster than Mo- bileNetV2 and ProxylessNAS-based solutions with 3.7-4.1× smaller peak SRAM. Our study suggests that the era of always-on tiny machine learning on IoT devices has arrived.

We focus on the task of generating sound from natural videos, and the sound should be both temporally and content-wise aligned with visual signals. This task is extremely challenging because some sounds generated outside a camera can not be inferred from video content. The model may be forced to learn an incorrect mapping between visual content and these irrelevant sounds. To address this challenge, we propose a framework named REGNET. In this framework, we first extract appearance and motion features from video frames to better distinguish the object that emits sound from complex background information. We then introduce an innovative audio forwarding regularizer that directly considers the real sound as input and outputs bottlenecked sound features. Using both visual and bottlenecked sound features for sound prediction during training provides stronger supervision for the sound prediction. The audio forwarding regularizer can control the irrelevant sound component and thus prevent the model from learning an incorrect mapping between video frames and sound emitted by the object that is out of the screen. During testing, the audio forwarding regularizer is removed to ensure that REGNET can produce purely aligned sound only from visual features. Extensive evaluations based on Amazon Mechanical Turk demonstrate that our method significantly improves both temporal and contentwise alignment. Remarkably, our generated sound can fool the human with a 68.12% success rate. Code and pre-trained models are publicly available at https://github.com/PeihaoChen/regnet.

Temporal action localization aims to accurately localize and recognize all possible action instances from an untrimmed video automatically. Most existing methods perform this task by first generating a set of proposals and then recognizing each independently. However, due to the complex structures and large content variations in action instances, recognizing them individually can be difficult. Fortunately, some proposals often share information regarding one specific action. Such information, which is ignored in existing methods, can be used to boost recognition performance. In this paper, we propose a novel mechanism, called relation attention, to exploit informative relations among proposals based on their appearance or optical flow features. Specifically, we propose a relation attention module to enhance representation power by capturing useful information from other proposals. This module does not change the dimensions of the original input and output and does not rely on any specific proposal generation methods or feature extraction backbone networks. Experimental results show that the proposed relation attention mechanism improves performance significantly on both Thumos14 and ActivityNet1.3 datasets compared to existing architectures. For example, relying on Structured Segment Networks (SSN), the proposed relation attention module helps to increase the mAP from 41.4 to 43.7 on the Thumos14 dataset and outperforms the state-of-the-art results.

The recent advances in deep learning have made it possible to generate photo-realistic images by using neural networks and even to extrapolate video frames from an input video clip. In this paper, for the sake of both furthering this exploration and our own interest in a realistic application, we study imageto-video translation and particularly focus on the videos of facial expressions. This problem challenges the deep neural networks by another temporal dimension comparing to the image-to-image translation. Moreover, its single input image fails most existing video generation methods that rely on recurrent models. We propose a user-controllable approach so as to generate video clips of various lengths from a single face image. The lengths and types of the expressions are controlled by users. To this end, we design a novel neural network architecture that can incorporate the user input into its skip connections and propose several improvements to the adversarial training method for the neural network. Experiments and user studies verify the effectiveness of our approach. Especially, we would like to highlight that even for the face images in the wild (downloaded from the Web and the authors’ own photos), our model can generate high-quality facial expression videos of which about 50% are labeled as real by Amazon Mechanical Turk workers.

Despite the success of deep learning for static image understanding, it remains unclear what are the most effective network architectures for spatial-temporal modeling in videos. In this paper, in contrast to the existing CNN+RNN or pure 3D convolution based approaches, we explore a novel spatialtemporal network (StNet) architecture for both local and global modeling in videos. Particularly, StNet stacks N successive video frames into a super-image which has 3N channels and applies 2D convolution on super-images to capture local spatial-temporal relationship. To model global spatialtemporal structure, we apply temporal convolution on the local spatial-temporal feature maps. Specifically, a novel temporal Xception block is proposed in StNet, which employs a separate channel-wise and temporal-wise convolution over the feature sequence of a video. Extensive experiments on the Kinetics dataset demonstrate that our framework outperforms several state-of-the-art approaches in action recognition and can strike a satisfying trade-off between recognition accuracy and model complexity. We further demonstrate the generalization performance of the leaned video representations on the UCF101 dataset.

Most of the recent progresses on visual question answering are based on recurrent neural networks (RNNs) with attention. Despite the success, these models are often timeconsuming and having difficulties in modeling long range dependencies due to the sequential nature of RNNs. We propose a new architecture, Positional Self-Attention with Coattention (PSAC), which does not require RNNs for video question answering. Specifically, inspired by the success of self-attention in machine translation task, we propose a Positional Self-Attention to calculate the response at each position by attending to all positions within the same sequence, and then add representations of absolute positions. Therefore, PSAC can exploit the global dependencies of question and temporal information in the video, and make the process of question and video encoding executed in parallel. Furthermore, in addition to attending to the video features relevant to the given questions (i.e., video attention), we utilize the co-attention mechanism by simultaneously modeling “what words to listen to” (question attention). To the best of our knowledge, this is the first work of replacing RNNs with selfattention for the task of visual question answering. Experimental results of four tasks on the benchmark dataset show that our model significantly outperforms the state-of-the-art on three tasks and attains comparable result on the Count task. Our model requires less computation time and achieves better performance compared with the RNNs-based methods. Additional ablation study demonstrates the effect of each component of our proposed model.

Segmenting objects in images and separating sound sources in audio are challenging tasks, in part because traditional approaches require large amounts of labeled data. In this paper we develop a neural network model for visual object segmentation and sound source separation that learns from natural videos through self-supervision. The model is an extension of recently proposed work that maps image pixels to sounds [1]. Here, we introduce a learning approach to disentangle concepts in the neural networks, and assign semantic categories to network feature channels to enable independent image segmentation and sound source separation after audio-visual training on videos. Our evaluations show that the disentangled model outperforms several baselines in semantic segmentation and sound source separation.

Humans are able to localize objects in the environment using both visual and auditory cues, integrating information from multiple modalities into a common reference frame. We introduce a system that can leverage unlabeled audiovisual data to learn to localize objects (moving vehicles) in a visual reference frame, purely using stereo sound at inference time. Since it is labor-intensive to manually annotate the correspondences between audio and object bounding boxes, we achieve this goal by using the co-occurrence of visual and audio streams in unlabeled videos as a form of self-supervision, without resorting to the collection of ground truth annotations. In particular, we propose a framework that consists of a vision ``teacher'' network and a stereo-sound ``student'' network. During training, knowledge embodied in a well-established visual vehicle detection model is transferred to the audio domain using unlabeled videos as a bridge. At test time, the stereo-sound student network can work independently to perform object localization using just stereo audio and camera meta-data, without any visual input. Experimental results on a newly collected Auditory Vehicles Tracking dataset verify that our proposed approach outperforms several baseline approaches. We also demonstrate that our cross-modal auditory localization approach can assist in the visual localization of moving vehicles under poor lighting conditions.

The explosive growth in video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive, making it expensive to deploy. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. Specifically, it can achieve the performance of 3D CNN but maintain 2D CNN's complexity. TSM shifts part of the channels along the temporal dimension; thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. We also extended TSM to online setting, which enables real-time low-latency online video recognition and video object detection. TSM is accurate and efficient: it ranks the first place on the Something-Something leaderboard upon publication; on Jetson Nano and Galaxy Note8, it achieves a low latency of 13ms and 35ms for online video recognition. The code is available at: https://github. com/mit-han-lab/temporal-shift-module.

Most state-of-the-art action localization systems process each action proposal individually, without explicitly exploiting their relations during learning. However, the relations between proposals actually play an important role in action localization, since a meaningful action always consists of multiple proposals in a video. In this paper, we propose to exploit the proposal-proposal relations using GraphConvolutional Networks (GCNs). First, we construct an action proposal graph, where each proposal is represented as a node and their relations between two proposals as an edge. Here, we use two types of relations, one for capturing the context information for each proposal and the other one for characterizing the correlations between distinct actions. Then we apply the GCNs over the graph to model the relations among different proposals and learn powerful representations for the action classification and localization. Experimental results show that our approach significantly outperforms the state-of-the-art on THUMOS14(49.1% versus 42.8%). Moreover, augmentation experiments on ActivityNet also verify the efficacy of modeling action proposal relationships.

Sounds originate from object motions and vibrations of surrounding air. Inspired by the fact that humans is capable of interpreting sound sources from how objects move visually, we propose a novel system that explicitly captures such motion cues for the task of sound localization and separation. Our system is composed of an end-to-end learnable model called Deep Dense Trajectory (DDT), and a curriculum learning scheme. It exploits the inherent coherence of audio-visual signals from a large quantities of unlabeled videos. Quantitative and qualitative evaluations show that comparing to previous models that rely on visual appearance cues, our motion based system improves performance in separating musical instrument sounds. Furthermore, it separates sound components from duets of the same category of instruments, a challenging problem that has not been addressed before.

Humans have a surprising capacity to induce general rules that describe the specific actions portrayed in a video sequence. The rules learned through this kind of process allow us to achieve similar goals to those shown in the video but in more general circumstances. Enabling an agent to achieve the same capacity represents a significant challenge. In this paper, we propose a Watch-Reason-Code(WRC) model to synthesise programs that describe the process carried out in a set of video sequences. The 'watch' stage is simply a video encoder that encodes videos to multiple feature vectors. The 'reason' stage takes as input the features from multiple diverse videos and generates a compact feature representation via a novel deviation-pooling method. The 'code' stage is a multi-sound decoder that the first step leverages to generate a draft program layout with possible useful statements and perceptions. Further steps then take these outputs and generate a fully structured, compile-able and executable program. We evaluate the effectiveness of our model in two video-to-program synthesis environments, Karel andVizDoom, showing that we can achieve the state-of-the-art under a variety of settings.

There has been a prominent emergence of work on video prediction, aiming to extrapolate the future video frames from the past. Existing temporal-based methods are limited to certain numbers of frames. In this paper, we study video prediction from a single still image in the facial expression domain, a.k.a, facial image-to-video translation. Our main approach, dubbed AffineGAN, associates each facial image with an expression intensity and leverages an affine transformation in the latent space. AffineGAN allows users to control the number of frames to predict as well as the expression intensity for each of them. Unlike previous intensity-based methods, We derive an inverse formulation to the affine transformation, enabling automatic inference of the facial expression intensities from videos --- manual annotation is not only tedious but also ambiguous as people express in various ways and have different opinions about the intensity of a facial image. Both quantitative and qualitative results verify the superiority of AffineGAN over the state of the arts. Notably, in a Turing test with web faces, more than 50% of the facial expression videos generated by AffineGAN are considered real by the Amazon Mechanical Turk workers. This work could improve users' communication experience by enabling them to conveniently and creatively produce expression GIFs, which are popular art forms in online messaging and social networks.

Humans reason with concepts and metaconcepts: we recognize red and blue from visual input; we also understand that they are colors, i.e., red is an instance of color. In this paper, we propose the visual concept-metaconcept learner (VCML) for joint learning of concepts and metaconcepts from images and associated question-answer pairs. The key is to exploit the bidirectional connection between visual concepts and metaconcepts. Visual representations provide grounding cues for predicting relations between unseen pairs of concepts. Knowing that red and blue are instances of color, we generalize to the fact that green is also an instance of color since they all categorize the hue of objects. Meanwhile, knowledge about metaconcepts empowers visual concept learning from limited, noisy, and even biased data. From just a few examples of purple cubes we can understand a new color purple, which resembles the hue of the cubes instead of the shape of them. Evaluation on both synthetic and real-world datasets validates our claims.

This paper studies Learning from Observations (LfO) for imitation learning with access to state-only demonstrations. In contrast to Learning from Demonstration (LfD) that involves both action and state supervisions, LfO is more practical in leveraging previously inapplicable resources (e.g., videos), yet more challenging due to the incomplete expert guidance. In this paper, we investigate LfO and its difference with LfD in both theoretical and practical perspectives. We first prove that the gap between LfD and LfO actually lies in the disagreement of inverse dynamics models between the imitator and expert, if following the modeling approach of GAIL. More importantly, the upper bound of this gap is revealed by a negative causal entropy which can be minimized in a model-free way. We term our method as Inverse-Dynamics-Disagreement-Minimization (IDDM) which enhances the conventional LfO method through further bridging the gap to LfD. Considerable empirical results on challenging benchmarks indicate that our method attains consistent improvements over other LfO counterparts.

Convolutional neural networks process input data by sending channel-wise feature response maps to subsequent layers. While a lot of progress has been made by making networks deeper, information from each channel can only be propagated from lower levels to higher levels in a hierarchical feed-forward manner. When viewing each filter in the convolutional layer as a neuron, those neurons are not communicating explicitly within each layer in CNNs. We introduce a novel network unit called Cross-channel Communication (C3) block, a simple yet effective module to encourage the neuron communication within the same layer. The C3 block enables neurons to exchange information through a micro neural network, which consists of a feature encoder, a message communicator, and a feature decoder, before sending the information to the next layer. With C3 block, each neuron accounts for the channel-wise responses from other neurons at the same layer and learns more discriminative and complementary representations. Extensive experiments for multiple computer vision tasks show that our proposed mechanism allows shallower networks to aggregate useful information within each layer, and performances outperform baseline deep networks and other competitive methods.

In this paper, we propose the novel principal backpropagation networks (PBNets) to revisit the backpropagation algorithms commonly used in training two-stream networks for video action recognition. We content that existing approaches always take all the frames/snippets for the backpropagation not optimal for video recognition since the desired actions only occur in a short period within a video. To remedy these drawbacks, we design a watch-and-choose mechanism. In particular, the watching stage exploits a dense snippet-wise temporal pooling strategy to discover the global characteristic for each input video, while the choosing phase only backpropagates a small number of representative snippets that are selected with two novel strategies, i.e., Max-rule and KL-rule. We prove that with the proposed selection strategies, performing the backpropagation on the selected subset is capable of decreasing the loss of the whole snippets as well. The proposed PBNets are evaluated on two standard video action recognition benchmarks UCF101 and HMDB51, where it surpasses the state of the arts consistently, but requiring less memory and computation to achieve high performance.

We address the challenging problem of weakly supervised temporal action localization from unconstrained web videos, where only the video-level action labels are available during training. Inspired by the adversarial erasing strategy in weakly supervised semantic segmentation, we propose a novel iterative-winners-out network. Specifically, we make two technical contributions: we propose an iterative training strategy, namely, winners-out, to select the most discriminative action instances in each training iteration and remove them in the next training iteration. This iterative process alleviates the "winner-takes-all" phenomenon that existing approaches tend to choose the video segments that strongly correspond to the video label but neglects other less discriminative video segments. With this strategy, our network is able to localize not only the most discriminative instances but also the less discriminative ones. To better select the target action instances in winners-out, we devise a class-discriminative localization technique. By employing the attention mechanism and the information learned from data, our technique is able to identify the most discriminative action instances effectively. The two key components are integrated into an end-to-end network to localize actions without using the frame-level annotations. Extensive experimental results demonstrate that our method outperforms the state-of-the-art weakly supervised approaches on ActivityNet1.3 and improves mAP from 16.9% to 20.5% on THUMOS14. Notably, even with weak video-level supervision, our method attains comparable accuracy to those employing frame-level supervisions.

Video-based action recognition with deep neural networks has shown remarkable progress. However, most of the existing approaches are too computationally expensive due to the complex network architecture. To address these problems, we propose a new real-time action recognition architecture, called Temporal Convolutional 3D Network (T-C3D), which learns video action representations in a hierarchical multi-granularity manner. Specifically, we combine a residual 3D convolutional neural network which captures complementary information on the appearance of a single frame and the motion between consecutive frames with a new temporal encoding method to explore the temporal dynamics of the whole video. Thus heavy calculations are avoided when doing the inference, which enables the method to be capable of real-time processing. On two challenging benchmark datasets, UCF101 and HMDB51, our method is significantly better than state-of-the-art real-time methods by over 5.4% in terms of accuracy and 2 times faster in terms of inference speed (969 frames per second), demonstrating comparable recognition performance to the state-of-the-art methods. The source code for the complete system as well as the pre-trained models are publicly available at https://github.com/tc3d.

The problem of video classification is inherently sequential and multimodal, and deep neural models hence need to capture and aggregate the most pertinent signals for a given input video. We propose Keyless Attention as an elegant and efficient means to more effectively account for the sequential nature of the data. Moreover, comparing a variety of multimodal fusion methods, we find that Multimodal Keyless Attention Fusion is the most successful at discerning interactions between modalities. We experiment on four highly heterogeneous datasets, UCF101, ActivityNet, Kinetics, and YouTube-8M to validate our conclusion, and show that our approach achieves highly competitive results. Especially on large-scale data, our method has great advantages in efficiency and performance. Most remarkably, our best single model can achieve 77.0% in terms of the top-1 accuracy and 93.2% in terms of the top-5 accuracy on the Kinetics validation set, and achieve 82.2% in terms of GAP@20 on the official YouTube-8M test set.

We study the problem of reconstructing an image from information stored at contour locations. We show that high-quality reconstructions with high fidelity to the source image can be obtained from sparse input, e.g., comprising less than 6% of image pixels. This is a significant improvement over existing contour-based reconstruction methods that require much denser input to capture subtle texture information and to ensure image quality. Our model, based on generative adversarial networks, synthesizes texture and details in regions where no input information is provided. The semantic knowledge encoded into our model and the sparsity of the input allows to use contours as an intuitive interface for semantically-aware image manipulation: local edits in contour domain translate to long-range and coherent changes in pixel space. We can perform complex structural changes such as changing facial expression by simple edits of contours. Our experiments demonstrate that humans as well as a face recognition system mostly cannot distinguish between our reconstructions and the source images.

Despite the recent success of end-to-end learned representations, hand-crafted optical flow features are still widely used in video analysis tasks. To fill this gap, we propose TVNet, a novel end-to-end trainable neural network, to learn optical-flow-like features from data. TVNet subsumes a specific optical flow solver, the TV-L1 method, and is initialized by unfolding its optimization iterations as neural layers. TVNet can therefore be used directly without any extra learning. Moreover, it can be naturally concatenated with other task-specific networks to formulate an end-to-end architecture, thus making our method more efficient than current multi-stage approaches by avoiding the need to pre-compute and store features on disk. Finally, the parameters of the TVNet can be further fine-tuned by end-to-end training. This enables TVNet to learn richer and task-specific patterns beyond exact optical flow. Extensive experiments on two action recognition benchmarks verify the effectiveness of the proposed approach. Our TVNet achieves better accuracies than all compared methods, while being competitive with the fastest counterpart in terms of features extraction time.

It is often laborious and costly to manually annotate videos for training high-quality video recognition models, so there has been some work and interest in exploring alternative, cheap, and yet often noisy and indirect training signals for learning the video representations. However, these signals are still coarse, supplying supervision at the whole video frame level, and subtle, sometimes enforcing the learning agent to solve problems that are even hard for humans. In this paper, we instead explore geometry, a grand new type of auxiliary supervision for the self-supervised learning of video representations. In particular, we extract pixel-wise geometry information as flow fields and disparity maps from synthetic imagery and real 3D movies, respectively. Although the geometry and high-level semantics are seemingly distant topics, surprisingly, we find that the convolutional neural networks pre-trained by the geometry cues can be effectively adapted to semantic video understanding tasks. In addition, we also find that a progressive training strategy can foster a better neural network for the video recognition task than blindly pooling the distinct sources of geometry cues together. Extensive results on video dynamic scene recognition and action recognition tasks show that our geometry guided networks significantly outperform the competing methods that are trained with other types of labeling-free supervision signals.

Recently, substantial research effort has focused on how to apply CNNs or RNNs to better capture temporal patterns in videos, so as to improve the accuracy of video classification. In this paper, however, we show that temporal information, especially longer-term patterns, may not be necessary to achieve competitive results on common trimmed video classification datasets. We investigate the potential of a purely attention based local feature integration. Accounting for the characteristics of such features in video classification, we propose a local feature integration framework based on attention clusters, and introduce a shifting operation to capture more diverse signals. We carefully analyze and compare the effect of different attention mechanisms, cluster sizes, and the use of the shifting operation, and also investigate the combination of attention clusters for multimodal integration. We demonstrate the effectiveness of our framework on three real-world video classification datasets. Our model achieves competitive results across all of these. In particular, on the large-scale Kinetics dataset, our framework obtains an excellent single model accuracy of 79.4% in terms of the top-1 and 94.0% in terms of the top-5 accuracy on the validation set.

Dense event captioning aims to detect and describe all events of interest contained in a video. Despite the advanced development in this area, existing methods tackle this task by making use of dense temporal annotations, which is dramatically source-consuming. This paper formulates a new problem: weakly supervised dense event captioning, which does not require temporal segment annotations for model training. Our solution is based on the one-to-one correspondence assumption, each caption describes one temporal segment, and each temporal segment has one caption, which holds in current benchmark datasets and most real world cases. We decompose the problem into a pair of dual problems: event captioning and sentence localization and present a cycle system to train our model. Extensive experimental results are provided to demonstrate the ability of our model on both dense event captioning and sentence localization in videos.

We marry two powerful ideas: deep representation learning for visual recognition and language understanding, and symbolic program execution for reasoning. Our neural-symbolic visual question answering (NS-VQA) system first recovers a structural scene representation from the image and a program trace from the question. It then executes the program on the scene representation to obtain an answer. Incorporating symbolic structure as prior knowledge offers three unique advantages. First, executing programs on a symbolic space is more robust to long program traces; our model can solve complex reasoning tasks better, achieving an accuracy of 99.8% on the CLEVR dataset. Second, the model is more data- and memory-efficient: it performs well after learning on a small number of training data; it can also encode an image into a compact representation, requiring less storage than existing methods for offline question answering. Third, symbolic program execution offers full transparency to the reasoning process; we are thus able to interpret and diagnose each execution step.

Video captioning has attracted an increasing amount of interest, due in part to its potential for improved accessibility and information retrieval. While existing methods rely on different kinds of visual features and model architectures, they do not make full use of pertinent semantic cues. We present a unified and extensible framework to jointly leverage multiple sorts of visual features and semantic attributes. Our novel architecture builds on LSTMs with two multi-faceted attention layers. These first learn to automatically select the most salient visual features or semantic attributes, and then yield overall representations for the input and output of the sentence generation component via custom feature scaling operations. Experimental results on the challenging MSVD and MSR-VTT datasets show that our framework outperforms previous work and performs robustly even in the presence of added noise to the features and attributes.

We address the problem of zero-shot event recognition in consumer videos. An event usually consists of multiple human-human and human-object interactions over a relative long period of time. A common approach proceeds by representing videos with banks of object and action concepts, but requires additional user inputs to specify the desired concepts per event. In this paper, we provide a fully automatic algorithm to select representative and reliable concepts for event queries. This is achieved by discovering event composition knowledge (DECK) from web images. To evaluate our proposed method, we use the standard zero-shot event detection protocol (ZeroMED), but also introduce a novel zero-shot event recounting (ZeroMER) problem to select supporting evidence of the events. Our ZeroMER formulation aims to select video snippets that are relevant and diverse. Evaluation on the challenging TRECVID MED dataset show that our proposed method achieves promising results on both tasks.

We propose a novel framework named StyleNet to address the task of generating attractive captions for images and videos with different styles. To this end, we devise a novel model component, named factored LSTM, which automatically distills the style factors in the monolingual text corpus. Then at runtime, we can explicitly control the style in the caption generation process so as to produce attractive visual captions with the desired style. Our approach achieves this goal by leveraging two sets of data: 1) factual image/video-caption paired data, and 2) stylized monolingual text data (e.g., romantic and humorous sentences). We show experimentally that StyleNet outperforms existing approaches for generating visual captions with different styles, measured in both automatic and human evaluation metrics on the newly collected FlickrStyle10K image caption dataset, which contains 10K Flickr images with corresponding humorous and romantic captions.

A Semantic Compositional Network (SCN) is developed for image captioning, in which semantic concepts (i.e., tags) are detected from the image, and the probability of each tag is used to compose the parameters in a long short-term memory (LSTM) network. The SCN extends each weight matrix of the LSTM to an ensemble of tag-dependent weight matrices. The degree to which each member of the ensemble is used to generate an image caption is tied to the image-dependent probability of the corresponding tag. In addition to captioning images, we also extend the SCN to generate captions for video clips. We qualitatively analyze semantic composition in SCNs, and quantitatively evaluate the algorithm on three benchmark datasets: COCO, Flickr30k, and Youtube2Text. Experimental results show that the proposed method significantly outperforms prior state-of-the-art approaches, across multiple evaluation metrics.

Rich and dense human labeled datasets are among the main enabling factors for the recent advance on visionlanguage understanding. Many seemingly distant annotations (e.g., semantic segmentation and visual question answering (VQA)) are inherently connected in that they reveal different levels and perspectives of human understandings about the same visual scenes — and even the same set of images (e.g., of COCO). The popularity of COCO correlates those annotations and tasks. Explicitly linking them up may significantly benefit both individual tasks and the unified vision and language modeling.,,We present the preliminary work of linking the instance segmentations provided by COCO to the questions and answers (QAs) in the VQA dataset, and name the collected links visual questions and segmentation answers (VQS). They transfer human supervision between the previously separate tasks, offer more effective leverage to existing problems, and also open the door for new research problems and models. We study two applications of the VQS data in this paper: supervised attention for VQA and a novel question-focused semantic segmentation task. For the former, we obtain state-of-the-art results on the VQA real multiple-choice task by simply augmenting the multilayer perceptrons with some attention features that are learned using the segmentation-QA links as explicit supervision. To put the latter in perspective, we study two plausible methods and compare them to an oracle method assuming that the instance segmentations are given at the test stage.

A natural image usually conveys rich semantic content and can be viewed from different angles. Existing image description methods are largely restricted by small sets of biased visual paragraph annotations, and fail to cover rich underlying semantics. In this paper, we investigate a semi-supervised paragraph generative framework that is able to synthesize diverse and semantically coherent paragraph descriptions by reasoning over local semantic regions and exploiting linguistic knowledge. The proposed Recurrent Topic-Transition Generative Adversarial Network (RTT-GAN) builds an adversarial framework between a structured paragraph generator and multi-level paragraph discriminators. The paragraph generator generates sentences recurrently by incorporating region-based visual and language attention mechanisms at each step. The quality of generated paragraph sentences is assessed by multi-level adversarial discriminators from two aspects, namely, plausibility at sentence level and topic-transition coherence at paragraph level. The joint adversarial training of RTT-GAN drives the model to generate realistic paragraphs with smooth logical transition between sentence topics. Extensive quantitative experiments on image and video paragraph datasets demonstrate the effectiveness of our RTT-GAN in both supervised and semi-supervised settings. Qualitative results on telling diverse stories for an image verify the interpretability of RTT-GAN.

Vast quantities of videos are now being captured at astonishing rates, but the majority of these are not labelled. To cope with such data, we consider the task of content-based activity recognition in videos without any manually labelled examples, also known as zero-shot video recognition. To achieve this, videos are represented in terms of detected visual concepts, which are then scored as relevant or irrelevant according to their similarity with a given textual query. In this paper, we propose a more robust approach for scoring concepts in order to alleviate many of the brittleness and low precision problems of previous work. Not only do we jointly consider semantic relatedness, visual reliability, and discriminative power. To handle noise and non-linearities in the ranking scores of the selected concepts, we propose a novel pairwise order matrix approach for score aggregation. Extensive experiments on the large-scale TRECVID Multimedia Event Detection data show the superiority of our approach.

Attributes possess appealing properties and benefit many computer vision problems, such as object recognition, learning with humans in the loop, and image retrieval. Whereas the existing work mainly pursues utilizing attributes for various computer vision problems, we contend that the most basic problem-how to accurately and robustly detect attributes from images-has been left under explored. Especially, the existing work rarely explicitly tackles the need that attribute detectors should generalize well across different categories, including those previously unseen. Noting that this is analogous to the objective of multi-source domain generalization, if we treat each category as a domain, we provide a novel perspective to attribute detection and propose to gear the techniques in multi-source domain generalization for the purpose of learning cross-category generalizable attribute detectors. We validate our understanding and approach with extensive experiments on four challenging datasets and three different problems.

Video concept learning often requires a large set oftraining samples. In practice, however, acquiring noise-free training labels with sufficient positive examples is very expensive. A plausible solution for training data collection is by sampling from the vast quantities of images and videos on the Web. Such a solution is motivated by the assumption that the retrieved images or videos are highly correlated with the query. Still, a number ofchallenges remain. First, Web videos are often untrimmed. Thus, only parts of the videos are relevant to the query. Second, the retrieved Web images are always highly relevant to the issued query. However, thoughtlessly utilizing the images in the video domain may even hurt the performance due to the well-known semantic drift and domain gap problems. As a result, a valid question is how Web images and videos interact for video concept learning. In this paper, we propose a Lead-Exceed Neural Network (LENN), which reinforces the training on Web images and videos in a curriculum manner. Specifically, the training proceeds by inputting frames of Web videos to obtain a network. The Web images are then filtered by the learnt network and the selected images are additionally fed into the network to enhance the architecture and further trim the videos. In addition, Long Short-Term Memory (LSTM) can be applied on the trimmed videos to explore temporal information. Encouraging results are reported on UCFIOl, TRECVID 2013 and 2014 MEDTest in the context ofboth action recognition and event detection. Without using human annotated exemplars, our proposed LENN can achieve 74.4% accuracy on UCFIOI dataset.

Automatically recognizing a large number of action categories from videos is of significant importance for video understanding. Most existing works focused on the design of more discriminative feature representation, and have achieved promising results when the positive samples are enough. However, very limited efforts were spent on recognizing a novel action without any positive exemplars, which is often the case in the real settings due to the large amount of action classes and the users' queries dramatic variations. To address this issue, we propose to perform action recognition when no positive exemplars of that class are provided, which is often known as the zero-shot learning. Different from other zero-shot learning approaches, which exploit attributes as the intermediate layer for the knowledge transfer, our main contribution is SIR, which directly leverages the semantic inter-class relationships between the known and unknown actions followed by label transfer learning. The inter-class semantic relationships are automatically measured by continuous word vectors, which learned by the skip-gram model using the large-scale text corpus. Extensive experiments on the UCF101 dataset validate the superiority of our method over fully-supervised approaches using few positive exemplars.

In this paper, we focus on complex event detection in internet videos while also providing the key evidences of the detection results. Convolutional Neural Networks (CNNs) have achieved promising performance in image classification and action recognition tasks. However, it remains an open problem how to use CNNs for video event detection and recounting, mainly due to the complexity and diversity of video events. In this work, we propose a flexible deep CNN infrastructure, namely Deep Event Network (DevNet), that simultaneously detects pre-defined events and provides key spatial-temporal evidences. Taking key frames of videos as input, we first detect the event of interest at the video level by aggregating the CNN features of the key frames. The pieces of evidences which recount the detection results, are also automatically localized, both temporally and spatially. The challenge is that we only have video level labels, while the key evidences usually take place at the frame levels. Based on the intrinsic property of CNNs, we first generate a spatial-temporal saliency map by back passing through DevNet, which then can be used to find the key frames which are most indicative to the event, as well as to localize the specific spatial position, usually an object, in the frame of the highly indicative area. Experiments on the large scale TRECVID 2014 MEDTest dataset demonstrate the promising performance of our method, both for event detection and evidence recounting.

Humans connect language and vision to perceive the world. How to build a similar connection for computers? One possible way is via visual concepts, which are text terms that relate to visually discriminative entities. We propose an automatic visual concept discovery algorithm using parallel text and visual corpora, it filters text terms based on the visual discriminative power of the associated images, and groups them into concepts using visual and semantic similarities. We illustrate the applications of the discovered concepts using bidirectional image and sentence retrieval task and image tagging task, and show that the discovered concepts not only outperform several large sets of manually selected concepts significantly, but also achieves the state-of-the-art performance in the retrieval task.

Contemporary video search and categorization are non-trivial tasks due to the massively increasing amount and content variety of videos. We put forward the study of visual saliency models in video. Such a model is employed to identify salient objects from the image background. Starting from the observation that motion information in video often attracts more human attention compared to static images, we devise a region contrast based saliency detection model using spatial-temporal cues (RCST). We introduce and study four saliency principles to realize the RCST. This generalizes the previous static image for saliency computational model to video. We conduct experiments on a publicly available video segmentation database where our method significantly outperforms seven state-of-the-art methods with respect to PR curve, ROC curve and visual comparison.