Xiwen Liang, Min Lin, Weiqi Ruan, Rongtao Xu, Yuecheng Liu, Jiaqi Chen, Bingqian Lin, Yuzheng Zhuang, Xiaodan Liang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Vision-language-action models have gained significant attention for their ability to model multimodal sequences in embodied instruction following tasks. However, most existing models rely on causal attention, which we find suboptimal for processing sequences composed of interleaved segments from different modalities. In this paper, we introduce Astra, a novel Transformer architecture featuring trajectory attention and learnable action queries, designed to efficiently process segmented multimodal trajectories and predict actions for imitation learning. Furthermore, we propose a contrastive dynamics learning objective to enhance the model’s understanding of environment dynamics and multimodal alignment, complementing the primary behavior cloning objective. Through extensive experiments on three large-scale robot manipulation benchmarks, Astra demonstrates substantial performance improvements over previous models.

Collecting data from simulated scenarios for training robotic skills provides a safer and more controllable alternative to real-world environments. However, it demands considerable effort, including the manual construction of simulation environments, the careful design of tasks, and the challenge of obtaining effective trajectories. These limitations hinder the efficiency of data collection from simulated scenarios. In this paper, we leverage the prior knowledge of Large Language Models (LLMs) and Large Multimodal Models (LMMs) to generate simulated scenarios and embodied tasks. We introduce a novel framework, ASCENT (Autonomous Skill learning toward Complex Embodied tasks with fouNdaTion models), designed to efficiently accomplish these tasks and generate trajectory data. ASCENT features a fully autonomous skill learning mechanism based on AI agent. During task training, the AI agent identifies suitable atomic skills from an atomic skill library to either directly complete the task or serve as an initial policy for further training. Newly acquired atomic skills are subsequently added to the library. To address training failures and enhance efficiency, the AI agent uses an LLM to automatically optimize the skill training process based on feedback received from simulations. Experimental results indicate that the number of training steps required for learning new tasks can be reduced by up to 65.9 %.

Recent advancements in Large Language Models (LLMs) have catalyzed numerous efforts to apply these technologies to embodied tasks, with a particular focus on high-level task planning and task decomposition. LLMs face challenges in understanding the physical world, especially regarding spatial, temporal, and causal relationships among objects and actions. Moreover, the current benchmarks for evaluating these relationships are limited. To further investigate this domain, we introduce a novel embodied task planning benchmark, ET-Plan-Bench. This benchmark features a controllable and diverse array of embodied tasks, varying in levels of difficulty and complexity. It is designed to evaluate two critical dimensions of LLMs’ application in embodied task understanding: spatial understanding (including relation constraints and occlusion of target objects) and temporal and causal comprehension of sequences of actions within an environment. Utilizing multi-source simulators as the backend simulator, ET-Plan-Bench provides immediate environmental feedback to LLMs, enabling dynamic interaction with the environment and the capacity for re-planning as necessary. We evaluated state-of-the-art open-source and closed-source foundational models, including GPT-4, Llama, and Mistral, using our proposed benchmark. While these models perform adequately on simple navigation tasks, their performance significantly deteriorates when con-fronted with tasks that demand a deeper understanding of spatial, temporal, and causal relationships. Consequently, our benchmark distinguishes itself as a large-scale, quantifiable, highly automated, and fine-grained diagnostic framework that presents a substantial challenge to the latest foundational models. We hope it will inspire and propel further research in embodied task planning utilizing foundational models. Code available at: https://github.com/ET-Plan-Bench/ET-Plan-Bench

Spatial Planning is a crucial part in the field of spatial intelligence, which requires the understanding and planning about object arrangements in space perspective. AI agents with the spatial planning ability can better adapt to various real-world applications, including robotic manipulation, automatic assembly, urban planning etc. Recent works have attempted to construct benchmarks for evaluating the spatial intelligence of Multimodal Large Language Models (MLLMs). Nevertheless, these benchmarks primarily focus on spatial reasoning based on typical Visual Question-Answering (VQA) forms, which suffers from the gap between abstract spatial understanding and concrete task execution. In this work, we take a step further to build a comprehensive benchmark called MineAnyBuild, aiming to evaluate the spatial planning ability of open-world AI agents in the Minecraft game. Specifically, MineAnyBuild requires an agent to generate executable architecture building plans based on the given multi-modal human instructions. It involves 4,000 curated spatial planning tasks and also provides a paradigm for infinitely expandable data collection by utilizing rich player-generated content. MineAnyBuild evaluates spatial planning through four core supporting dimensions: spatial understanding, spatial reasoning, creativity, and spatial commonsense. Based on MineAnyBuild, we perform a comprehensive evaluation for existing MLLM-based agents, revealing the severe limitations but enormous potential in their spatial planning abilities. We believe our MineAnyBuild will open new avenues for the evaluation of spatial intelligence and help promote further development for open-world AI agents capable of spatial planning.

In this article, we propose a distributional policy-gradient method based on distributional reinforcement learning (RL) and policy gradient. Conventional RL algorithms typically estimate the expectation of return, given a state-action pair. Furthermore, distributional RL algorithms consider the return as a random variable and estimate the return distribution that can characterize the probability of different returns resulted by environmental uncertainties. Thus, the return distribution provides more valuable information than its expectation, leading to superior policies in general. Although distributional RL has been investigated widely in value-based RL methods, very few policy-gradient methods take advantage of distributional RL. To bridge this research gap, we propose a distributional policy-gradient method by introducing a distributional value function to the policy gradient (DVDPG). We estimate the distribution of policy gradient instead of the expectation estimated in conventional policy-gradient RL methods. Furthermore, we propose two policy-gradient value sampling mechanisms to do policy improvement. First, we propose a distribution-probability-sampling method that samples the policy-gradient value according to the quantile probability of return distribution. Second, a uniform sample mechanism is proposed. With our sample mechanisms, the proposed distributional policy-gradient method enhances the stochasticity of the policy gradient, improving the exploration efficiency and benefiting to avoid falling into local optimal solutions. In sparse-reward tasks, the distribution-probability-sampling method outperforms the uniform sample mechanism. In dense-reward tasks, the two sample mechanisms perform similarly. Moreover, we show that the conventional policy-gradient method is a special case of the proposed method. Experimental results on various sparse-reward and dense-reward OpenAI-gym tasks illustrate the efficiency of the proposed method, outperforming baselines in almost environments.

Although visual navigation has been extensively studied using deep reinforcement learning, online learning for real-world robots remains a challenging task. Recent work directly learned from offline dataset to achieve broader generalization in the real-world tasks, which, however, faces the out-of-distribution (OOD) issue and potential robot localization failures in a given map for unseen observation. This significantly drops the success rates and even induces collision. In this paper, we present a self-correcting visual navigation method, SCALE, that can autonomously prevent the robot from the OOD situations without human intervention. Specifically, we develop an image-goal conditioned offline reinforcement learning method based on implicit Q-learning (IQL). When facing OOD observation, our novel localization recovery method generates the potential future trajectories by learning from the navigation affordance, and estimates the future novelty via random network distillation (RND). A tailored cost function searches for the candidates with the least novelty that can lead the robot to the familiar places. We collect offline data and conduct evaluation experiments in three real-world urban scenarios. Experiment results show that SCALE outperforms the previous state-of-the-art methods for open-world navigation with a unique capability of localization recovery, significantly reducing the need for human intervention. Code is available at https://github.com/KubeEdge4Robotics/ScaleNav.

High-quality traffic flow generation is the core module in building simulators for autonomous driving. However, the majority of available simulators are incapable of replicating traffic patterns that accurately reflect the various features of real-world data while also simulating human-like reactive responses to the tested autopilot driving strategies. Taking one step forward to addressing such a problem, we propose Realistic Interactive TrAffic flow (RITA) as an integrated component of existing driving simulators to provide high-quality traffic flow for the evaluation and optimization of the tested driving strategies. RITA is developed with consideration of three key features, i.e., fidelity, diversity, and controllability, and consists of two core modules called RITABackend and RITAKit. RITABackend is built to support vehicle-wise control and provide traffic generation models from real-world datasets, while RITAKit is developed with easy-to-use interfaces for controllable traffic generation via RITABackend. We demonstrate RITA’s capacity to create diversified and high-fidelity traffic simulations in several highly interactive highway scenarios. The experimental findings demonstrate that our produced RITA traffic flows exhibit all three key features, hence enhancing the completeness of driving strategy evaluation. Moreover, we showcase the possibility for further improvement of baseline strategies through online fine-tuning with RITA traffic flows.