In this paper, we propose an efficient multi-level convolution architecture for 3D visual grounding. Conventional methods are difficult to meet the requirements of real-time inference due to the two-stage or point-based architecture. Inspired by the success of multi-level fully sparse convolutional architecture in 3D object detection, we aim to build a new 3D visual grounding framework following this technical route. However, as in 3D visual grounding task the 3D scene representation should be deeply interacted with text features, sparse convolution-based architecture is inefficient for this interaction due to the large amount of voxel features. To this end, we propose text-guided pruning (TGP) and completion-based addition (CBA) to deeply fuse 3D scene representation and text features in an efficient way by gradual region pruning and target completion. Specifically, TGP iteratively sparsifies the 3D scene representation and thus efficiently interacts the voxel features with text features by cross-attention. To mitigate the affect of pruning on delicate geometric information, CBA adaptively fixes the over-pruned region by voxel completion with negligible computational overhead. Compared with previous single-stage methods, our method achieves top inference speed and surpasses previous fastest method by 100% FPS. Our method also achieves state-of-the-art accuracy even compared with two-stage methods, with +1.13 lead of [email protected] on ScanRefer, and +2.6 and +3.2 leads on NR3D and SR3D respectively. The code is available at https://github.com/GWxuan/TSP3D.

While multimodal large language models demonstrate strong performance in complex reasoning tasks, they pose significant challenges related to model complexity during deployment, especially for resource-limited devices. In this paper, we propose an automatic pruning method for large vision-language models to enhance the efficiency of multimodal reasoning. Conventional methods rely on the training data of the original model to select the proper pruning ratio for different network components. However, these methods are impractical for large vision-language models due to the unaffordable search costs caused by web-scale training corpus. In contrast, our approach only leverages a small number of samples to search for the desired pruning policy by maximizing its generalization ability on unknown training data while maintaining the model accuracy, which enables the achievement of an optimal trade-off between accuracy and efficiency for large visual language models. Specifically, we formulate the generalization gap of the pruning strategy using the structural risk minimization principle. Based on both task performance and generalization capability, we iteratively search for the optimal pruning policy within a given search space and optimize the vision projector to evolve the search space with higher upper bound of performance. We conduct extensive experiments on the ScienceQA, Vizwiz, MM-vet, and LLaVA-Bench datasets for the task of visual question answering. Using only 64 samples for pruning policy search, EfficientLLaVA achieves an accuracy of 83.05% on ScienceQA, along with a ×1.8 speedup compared to the dense LLaVA-v1.5-7B model. The code is available at: https://github.com/liangynTHU/EfficientLLaVA

Mobile manipulation is the fundamental challenge for robotics to assist humans with diverse tasks and environments in everyday life. However, conventional mobile manipulation approaches often struggle to generalize across different tasks and environments because of the lack of large-scale training. In contrast, recent advances in vision-language-action (VLA) models have shown impressive generalization capabilities, but these foundation models are developed for fixed-base manipulation tasks. Therefore, we propose an efficient policy adaptation framework named MoManipVLA to transfer pre-trained VLA models of fix-base manipulation to mobile manipulation, so that high generalization ability across tasks and environments can be achieved in mobile manipulation policy. Specifically, we utilize pre-trained VLA models to generate waypoints of the end-effector with high generalization ability. We design motion planning objectives for the mobile base and the robot arm, which aim at maximizing the physical feasibility of the trajectory. Finally, we present an efficient bi-level objective optimization framework for trajectory generation, where the upper-level optimization predicts way-points for base movement to enhance the manipulator policy space, and the lower-level optimization selects the optimal end-effector trajectory to complete the manipulation task. Extensive experimental results on OVMM and the real world demonstrate that MoManipVLA achieves a 4.2% higher success rate than the state-of-the-art mobile manipulation, and only requires 50 training cost for real world deployment due to the strong generalization ability in the pre-trained VLA models. Our project page can be found here.

In this paper, we propose a general framework for universal zero-shot goal-oriented navigation. Existing zero-shot methods build inference framework upon large language models (LLM) for specific tasks, which differs a lot in overall pipeline and fails to generalize across different types of goal. Towards the aim of universal zero-shot navigation, we propose a uniform graph representation to unify different goals, including object category, instance image and text description. We also convert the observation of agent into an online maintained scene graph. With this consistent scene and goal representation, we preserve most structural information compared with pure text and are able to leverage LLM for explicit graph-based reasoning. Specifically, we conduct graph matching between the scene graph and goal graph at each time instant and propose different strategies to generate long-term goal of exploration according to different matching states. The agent first iteratively searches subgraph of goal when zero-matched. With partial matching, the agent then utilizes coordinate projection and anchor pair alignment to infer the goal location. Finally scene graph correction and goal verification are applied for perfect matching. We also present a blacklist mechanism to enable robust switch between stages. Extensive experiments on several benchmarks show that our UniGoal achieves state-of-the-art zero-shot performance on three studied navigation tasks with a single model, even outperforming task-specific zero-shot methods and supervised universal methods. Project Page.

Visual navigation with an image as goal is a fundamental and challenging problem. Conventional methods either rely on end-to-end RL learning or modular-based policy with topological graph or BEV map as memory, which cannot fully model the geometric relationship between the explored 3D environment and the goal image. In order to efficiently and accurately localize the goal image in 3D space, we build our navigation system upon the renderable 3D gaussian (3DGS) representation. However, due to the computational intensity of 3DGS optimization and the large search space of 6-DoF camera pose, directly leveraging 3DGS for image localization during agent exploration process is prohibitively inefficient. To this end, we propose IGL-Nav, an Incremental 3D Gaussian Localization framework for efficient and 3D-aware image-goal navigation. Specifically, we incrementally update the scene representation as new images arrive with feed-forward monocular prediction. Then we coarsely localize the goal by leveraging the geometric information for discrete space matching, which can be equivalent to efficient 3D convolution. When the agent is close to the goal, we finally solve the fine target pose with optimization via differentiable rendering. The proposed IGL-Nav outperforms existing state-of-the-art methods by a large margin across diverse experimental configurations. It can also handle the more challenging free-view image-goal setting and be deployed on real-world robotic platform using a cellphone to capture goal image at arbitrary pose. Project page: https://gwxuan.github.io/IGL-Nav/.

Embodied tasks require the agent to fully understand 3D scenes simultaneously with its exploration, so an online, real-time, fine-grained and highly-generalized 3D perception model is desperately needed. Since high-quality 3D data is limited, directly training such a model in 3D is infeasible. Meanwhile, vision foundation models (VFM) has revolutionized the field of 2D computer vision with superior performance, which makes the use of VFM to assist embodied 3D perception a promising direction. However, most existing VFM-assisted 3D perception methods are either offline or too slow that cannot be applied in practical embodied tasks. In this paper, we aim to leverage Segment Anything Model (SAM) for real-time 3D instance segmentation in an online setting. This is a challenging problem since future frames are not available in the input streaming RGB-D video, and an instance may be observed in several frames so efficient object matching between frames is required. To address these challenges, we first propose a geometric-aware query lifting module to represent the 2D masks generated by SAM by 3D-aware queries, which is then iteratively refined by a dual-level query decoder. In this way, the 2D masks are transferred to fine-grained shapes on 3D point clouds. Benefit from the query representation for 3D masks, we can compute the similarity matrix between the 3D masks from different views by efficient matrix operation, which enables real-time inference. Experiments on ScanNet, ScanNet200, SceneNN and 3RScan show our method achieves state-of-the-art performance among online 3D perception models, even outperforming offline VFM-assisted 3D instance segmentation methods by a large margin. Our method also demonstrates great generalization ability in several zero-shot dataset transferring experiments and show great potential in data-efficient setting.

Recent trends in SLAM and visual navigation have embraced 3D Gaussians as the preferred scene representation, highlighting the importance of estimating camera poses from a single image using a pre-built Gaussian model. However, existing approaches typically rely on an iterative render-compare-refine loop, where candidate views are first rendered using NeRF or Gaussian Splatting, then compared against the target image, and finally, discrepancies are used to update the pose. This multi-round process incurs significant computational overhead, hindering real-time performance in robotics. In this paper, we propose iGaussian, a two-stage feed-forward framework that achieves real-time camera pose estimation through direct 3D Gaussian inversion. Our method first regresses a coarse 6DoF pose using a Gaussian Scene Prior-based Pose Regression Network with spatial uniform sampling and guided attention mechanisms, then refines it through feature matching and multi-model fusion. The key contribution lies in our cross-correlation module that aligns image embeddings with 3D Gaussian attributes without differentiable rendering, coupled with a Weighted Multiview Predictor that fuses features from Multiple strategically sampled viewpoints. Experimental results on the NeRF Synthetic, Mip-NeRF 360, and T&T+DB datasets demonstrate a significant performance improvement over previous methods, reducing median rotation errors to 0.2° while achieving 2.87 FPS tracking on mobile robots, which is an impressive 10× speedup compared to optimization-based approaches. Project page: https://github.com/pythongod-exe/iGaussian

Enabling embodied agents to complete complex human instructions from natural language is crucial to autonomous systems in household services. Conventional methods can only accomplish human instructions in the known environment where all interactive objects are provided to the embodied agent, and directly deploying the existing approaches for the unknown environment usually generates infeasible plans that manipulate non-existing objects. On the contrary, we propose an embodied instruction following (EIF) method for complex tasks in the unknown environment, where the agent efficiently explores the unknown environment to generate feasible plans with existing objects to accomplish abstract instructions. Specifically, we build a hierarchical embodied instruction following framework including the high-level task planner and the low-level exploration controller with multimodal large language models. We then construct a semantic representation map of the scene with dynamic region attention to demonstrate the known visual clues, where the goal of task planning and scene exploration is aligned for human instruction. For the task planner, we generate the feasible step-by-step plans for human goal accomplishment according to the task completion process and the known visual clues. For the exploration controller, the optimal navigation or object interaction policy is predicted based on the generated step-wise plans and the known visual clues. The experimental results demonstrate that our method can achieve 45.09% success rate in 204 complex human instructions such as making breakfast and tidying rooms in large house-level scenes. Code and supplementary are available at https://gary3410.github.io/eif_unknown/.

In this paper, we propose a novel network framework for indoor 3D object detection to handle variable input frame numbers in practical scenarios. Existing methods only consider fixed frames of input data for a single detector, such as monocular RGB-D images or point clouds reconstructed from dense multi-view RGB-D images. While in practical application scenes such as robot navigation and manipulation, the raw input to the 3D detectors is the RGB-D images with variable frame numbers instead of the reconstructed scene point cloud. However, the previous approaches can only handle fixed frame input data and have poor performance with variable frame input. In order to facilitate 3D object detection methods suitable for practical tasks, we present a novel 3D detection framework named AnyView for our practical applications, which generalizes well across different numbers of input frames with a single model. To be specific, we propose a geometric learner to mine the local geometric features of each input RGB-D image frame and implement local-global feature interaction through a designed spatial mixture module. Meanwhile, we further utilize a dynamic token strategy to adaptively adjust the number of extracted features for each frame, which ensures consistent global feature density and further enhances the generalization after fusion. Extensive experiments on the ScanNet dataset show our method achieves both great generalizability and high detection accuracy with a simple and clean architecture containing a similar amount of parameters with the baselines.

In this paper, we propose an autonomous robot packing system named RoboPacker designed to tightly store cluttered general objects into shipping boxes with high space utilization, which is a fundamental process in numerous industrial applications. However, achieving tight packaging for general objects often demands significant labor from human packers, particularly in high-throughput scenes. Compared to existing robot packing approaches, RoboPacker effectively overcomes challenges such as diverse object appearances, severe occlusion, and crowded packing spaces. Specifically, we propose an open-vocabulary shape estimation method to reconstruct complete point clouds for cluttered objects. We also design effective interactions with object clutter to gather informative visual clues for shape estimation under high uncertainty. Additionally, we introduce a hierarchical reinforcement learning framework to optimize packing order, location, and orientation for maximum space utilization. The robotic packing system integrates these techniques with feasible manipulation methods for real-world implementation. In this way, RoboPacker achieves efficient packing of novel and irregular objects, which is more suitable for real deployment environments. The Real-world experiments demonstrate RoboPacker can tightly pack 20 densely cluttered everyday objects from 8 seen and 4 novel classes into the 40×40×20 cm shipping box with a 73.3% success rate. The demonstration video can be found at https://gary3410.github.io/RoboPacker/.

In this paper, we propose an accurate post-training quantization framework of diffusion models (APQ-DM) for efficient image generation. Conventional quantization frameworks learn shared quantization functions for tensor discretization regardless of the generation timesteps in diffusion models, while the activation distribution differs significantly across various timesteps. Meanwhile, the calibration images are acquired in random timesteps which fail to provide sufficient information for generalizable quantization function learning. Both issues cause sizable quantization errors with obvious image generation performance degradation. On the contrary, we design distribution-aware quantization functions for activation discretization in different timesteps and search the optimal timesteps for informative calibration image generation, so that our quantized diffusion model can reduce the discretization errors with negligible computational overhead. Specifically, we partition various timestep quantization functions into different groups according to the importance weights, which are optimized by differentiable search algorithms. We also extend structural risk minimization principle for informative calibration image generation to enhance the generalization ability in the deployment of quantized diffusion model. Extensive experimental results show that our method outperforms the state-of-the-art post-training quantization of diffusion model by a sizable margin with similar computational cost11Code is available at https://github.com/ChangyuanWang17/APQ-DM.

In this paper, we propose a new framework for online 3D scene perception. Conventional 3D scene perception methods are offline, i.e., take an already reconstructed 3D scene geometry as input, which is not applicable in robotic applications where the input data is streaming RGB-D videos rather than a complete 3D scene reconstructed from precollected RGB-D videos. To deal with online 3D scene perception tasks where data collection and perception should be performed simultaneously, the model should be able to process 3D scenes frame by frame and make use of the temporal information. To this end, we propose an adapter-based plug-and-play module for the backbone of 3D scene perception model, which constructs memory to cache and aggregate the extracted RGB-D features to empower offline models with temporal learning ability. Specifically, we propose a queued memory mechanism to cache the supporting point cloud and image features. Then we devise aggregation modules which directly perform on the memory and pass temporal information to current frame. We further propose 3D-to-2D adapter to enhance image features with strong global context. Our adapters can be easily inserted into mainstream offline architectures of different tasks and significantly boost their performance on online tasks. Extensive experiments on ScanNet and SceneNN datasets demonstrate our approach achieves leading performance on three 3D scene perception tasks compared with state-of-the-art online methods by simply finetuning existing offline models, without any model and task-specific designs. Project page.

In this paper, we present a tensor decomposition and low-rank recovery approach (LowRankOcc) for vision-based 3D semantic occupancy prediction. Conventional methods model outdoor scenes with fine-grained 3D grids, but the sparsity of non-empty voxels introduces consider-able spatial redundancy, leading to potential overfitting risks. In contrast, our approach leverages the intrinsic low-rank property of 3D occupancy data, factorizing voxel representations into low-rank components to efficiently mitigate spatial redundancy without sacrificing performance. Specifically, we present the Vertical-Horizontal (VH) de-composition block factorizes 3D tensors into vertical vectors and horizontal matrices. With our “decomposition-encoding-recovery” framework, we encode 3D contexts with only 1/2D convolutions and poolings, and subsequently recover the encoded compact yet informative context features back to voxel representations. Experimental results demonstrate that LowRankOcc achieves state-of-the-art performances in semantic scene completion on the Se-manticKITTI dataset and 3D occupancy prediction on the nuScenes dataset.

In this paper, we propose a post-training quantization framework of large vision-language models (LVLMs) for efficient multi-modal inference. Conventional quantization methods sequentially search the layer-wise rounding functions by minimizing activation discretization errors, which fails to acquire optimal quantization strategy without considering cross-layer dependency. On the contrary, we mine the cross-layer dependency that significantly influences discretization errors of the entire vision-language model, and embed this dependency into optimal quantization strategy searching with low search cost. Specifically, we observe the strong correlation between the activation entropy and the cross-layer dependency concerning output discretization errors. Therefore, we employ the entropy as the proxy to partition blocks optimally, which aims to achieve satisfying trade-offs between discretization errors and the search cost. Moreover, we optimize the visual encoder to disentangle the cross-layer dependency for fine-grained decomposition of search space, so that the search cost is further reduced without harming the quantization accuracy. Experimental results demonstrate that our method compresses the memory by 2.78x and increase generate speed by 1.44x about 13B LLaVA model without performance degradation on diverse multi-modal reasoning tasks.

In this paper, we propose a new framework for zero-shot object navigation. Existing zero-shot object navigation methods prompt LLM with the text of spatially closed objects, which lacks enough scene context for in-depth reasoning. To better preserve the information of environment and fully exploit the reasoning ability of LLM, we propose to represent the observed scene with 3D scene graph. The scene graph encodes the relationships between objects, groups and rooms with a LLM-friendly structure, for which we design a hierarchical chain-of-thought prompt to help LLM reason the goal location according to scene context by traversing the nodes and edges. Moreover, benefit from the scene graph representation, we further design a re-perception mechanism to empower the object navigation framework with the ability to correct perception error. We conduct extensive experiments on MP3D, HM3D and RoboTHOR environments, where SG-Nav surpasses previous state-of-the-art zero-shot methods by more than \textbf{10\%} SR on all benchmarks, while the decision process is explainable. To the best of our knowledge, SG-Nav is the first zero-shot method that achieves even higher performance than supervised object navigation methods on the challenging MP3D benchmark. Code of this project will be released in the final version.

In this paper, we propose a weakly-supervised approach for 3D object detection, which makes it possible to train a strong 3D detector with position-level annotations (i.e. annotations of object centers and categories). In order to remedy the information loss from box annotations to centers, our method makes use of synthetic 3D shapes to convert the position-level annotations into virtual scenes with box-level annotations, and in turn utilizes the fully-annotated virtual scenes to complement the real labels. Specifically, we first present a shape-guided label-enhancement method, which assembles 3D shapes into physically reasonable virtual scenes according to the coarse scene layout extracted from position-level annotations. Then we transfer the information contained in the virtual scenes back to real ones by applying a virtual-to-real domain adaptation method, which refines the annotated object centers and additionally supervises the training of detector with the virtual scenes. Since the shape-guided label enhancement method generates virtual scenes by human-heuristic physical constraints, the layout of the fixed virtual scenes may be unreasonable with varied object combinations. To address this, we further present differentiable label enhancement to optimize the virtual scenes including object scales, orientations and locations in a data-driven manner. Moreover, we further propose a label-assisted self-training strategy to fully exploit the capability of detector. By reusing the position-level annotations and virtual scenes, we fuse the information from both domains and generate box-level pseudo labels on the real scenes, which enables us to directly train a detector in fully-supervised manner. Extensive experiments on the widely used ScanNet and Matterport3D datasets show that our approach surpasses current weakly-supervised and semi-supervised methods by a large margin, and achieves comparable detection performance with some popular fully-supervised methods with less than 5% of the labeling labor.

In this paper, we propose binary sparse convolutional networks called BSC-Net for efficient point cloud analysis. We empirically observe that sparse convolution operation causes larger quantization errors than standard convolution. However, conventional network quantization methods directly binarize the weights and activations in sparse convolution, resulting in performance drop due to the significant quantization loss. On the contrary, we search the optimal subset of convolution operation that activates the sparse convolution at various locations for quantization error alleviation, and the performance gap between real-valued and binary sparse convolutional networks is closed without complexity overhead. Specifically, we first present the shifted sparse convolution that fuses the information in the receptive field for the active sites that match the pre-defined positions. Then we employ the differentiable search strategies to discover the optimal opsitions for active site matching in the shifted sparse convolution, and the quantization errors are significantly alleviated for efficient point cloud analysis. For fair evaluation of the proposed method, we empirically select the recently advances that are beneficial for sparse convolution network binarization to construct a strong baseline. The experimental results on ScanNet and NYU Depth v2 show that our BSC-Net achieves significant improvement upon our srtong baseline and outperforms the state-of-the-art network binarization methods by a remarkable margin without additional computation overhead for binarizing sparse convolutional networks.

Due to the high price and heavy energy consumption of GPUs, deploying deep models on IoT devices such as microcontrollers makes significant contributions for ecological AI. Conventional methods successfully enable convolutional neural network inference of high resolution images on microcontrollers, while the framework for vision transformers that achieve the state-of-the-art performance in many vision applications still remains unexplored. In this paper, we propose a hardware-algorithm co-optimizations method called MCUFormer to deploy vision transformers on microcontrollers with extremely limited memory, where we jointly design transformer architecture and construct the inference operator library to fit the memory resource constraint. More specifically, we generalize the one-shot network architecture search (NAS) to discover the optimal architecture with highest task performance given the memory budget from the microcontrollers, where we enlarge the existing search space of vision transformers by considering the low-rank decomposition dimensions and patch resolution for memory reduction. For the construction of the inference operator library of vision transformers, we schedule the memory buffer during inference through operator integration, patch embedding decomposition, and token overwriting, allowing the memory buffer to be fully utilized to adapt to the forward pass of the vision transformer. Experimental results demonstrate that our MCUFormer achieves 73.62\% top-1 accuracy on ImageNet for image classification with 320KB memory on STM32F746 microcontroller. Code is available at https://github.com/liangyn22/MCUFormer.

In this article, we propose extremely low-precision vision transformers called Quantformer for efficient inference. Conventional network quantization methods directly quantize weights and activations of fully-connected layers without considering properties of transformer architectures. Quantization sizably deviates the self-attention compared with full-precision counterparts, and the shared quantization strategy for diversely distributed patch features causes severe quantization errors. To address these issues, we enforce the self-attention rank in quantized transformers to mimic that in full-precision counterparts with capacity-aware distribution for information retention, and quantize patch features with group-wise discretization strategy for quantization error minimization. Specifically, we efficiently preserve the self-attention rank consistency by minimizing the distance between the self-attention in quantized and real-valued transformers with adaptive concentration degree, where the optimal concentration degree is selected according to the self-attention entropy for model capacity adaptation. Moreover, we partition patch features in different dimensions with differentiable group assignment, so that features in different groups leverage various discretization strategies with minimal rounding and clipping errors. Experimental results show that our Quantformer outperforms the state-of-the-art network quantization methods by a sizable margin across various vision transformer architectures in image classification and object detection. We also integrate our Quantformer with mixed-precision quantization to further enhance the performance of the vanilla models.

In this paper, we propose a weakly-supervised approach for 3D object detection, which makes it possible to train a strong 3D detector with position-level annotations (i.e. annotations of object centers). In order to remedy the information loss from box annotations to centers, our method, namely Back to Reality (BR), makes use of synthetic 3D shapes to convert the weak labels into fully-annotated virtual scenes as stronger supervision, and in turn utilizes the perfect virtual labels to complement and refine the real labels. Specifically, we first assemble 3D shapes into physically reasonable virtual scenes according to the coarse scene layout extracted from position-level annotations. Then we go back to reality by applying a virtual-to-real domain adaptation method, which refine the weak labels and additionally supervise the training of detector with the virtual scenes. Furthermore, we propose a more challenging benckmark for indoor 3D object detection with more diversity in object sizes for better evaluation. With less than 5% of the labeling labor, we achieve comparable detection performance with some popular fully-supervised approaches on the widely used ScanNet dataset. Code is available at: https://github.com/wyf-ACCEPT/BackToReality.

In this paper, we propose an uncertainty-aware multi-resolution learning for point cloud segmentation, named PointRas. Most existing works for point cloud segmentation design encoder networks to obtain better representation of local space in point cloud. However, few of them investigate the utilization of features in the lower resolutions produced by encoders and consider the contextual learning between various resolutions in decoder network. To address this, we propose to utilize the descriptive characteristic of point clouds in the lower resolutions. Taking reference to core steps of rasterization in 2D graphics where the properties of pixels in high density are interpolated from a few primitive shapes in rasterization rendering, we use the similar strategy where prediction maps in lower resolution are iteratively regressed and upsampled into higher resolutions. Moreover, to remedy the potential information deficiency of lower-resolution point cloud, we refine the predictions in each resolution under the criterion of uncertainty selection, which notably enhances the representation ability of the point cloud in lower resolutions. Our proposed PointRas module can be incorporated into the backbones of various point cloud segmentation frameworks, and brings only marginal computational cost. We evaluate the proposed method on challenging datasets including ScanNet, S3DIS, NPM3D, STPLS3D and ScanObjectNN, and consistently improve the performance in comparison with the state-of-the-art methods.