CoRL 2025 论文列表

Modeling the nuanced, multimodal nature of human driving remains a core challenge for autonomous systems, as existing methods often fail to capture the diversity of plausible behaviors in complex real-world scenarios. In this work, we introduce a novel benchmark and end-to-end planner for modeling realistic multimodality in autonomous driving decisions. We propose a Gaussian Mixture Model (GMM)-based diffusion model designed to explicitly capture human-like, multimodal driving decisions in diverse contexts. Our model achieves state-of-the-art performance on current benchmarks, but reveals weaknesses in standard evaluation practices, which rely on single ground-truth trajectories or coarse closed-loop metrics while often penalizing diverse yet plausible alternatives. To address this limitation, we further develop a human-in-the-loop simulation benchmark that enables finer-grained evaluations and measures multimodal realism in challenging driving settings. Our code, models, and benchmark data will be released to promote more accurate and human-aware evaluation of autonomous driving models.

Humans naturally exhibit bilateral symmetry in their gross manipulation skills, effortlessly mirroring simple actions between left and right hands. Bimanual robots—which also feature bilateral symmetry—should similarly exploit this property to perform tasks with either hand. Unlike humans, who often favor a dominant hand for fine dexterous skills, robots should ideally execute ambidextrous manipulation with equal proficiency. To this end, we introduce SYMDEX (SYMmetric DEXterity), a reinforcement learning framework for ambidextrous bi-manipulation that leverages the robot's inherent bilateral symmetry as an inductive bias. SYMDEX decomposes complex bimanual manipulation tasks into per-hand subtasks and trains dedicated policies for each. By exploiting bilateral symmetry via equivariant neural networks, experience from one arm is inherently leveraged by the opposite arm. We then distill the subtask policies into a global ambidextrous policy that is independent of the hand-task assignment. We evaluate SYMDEX on six challenging simulated manipulation tasks and demonstrate successful real-world deployment on two of them. Our approach outperforms baselines on more complex, asymmetric tasks, where the left and right hands perform different roles. We further demonstrate SYMDEX’s scalability by extending it to a four-arm manipulation setup, where our symmetry-aware policies enable effective multi-arm collaboration and coordination. Our results highlight how structural symmetry as inductive bias in policy learning enhances sample efficiency, robustness, and generalization across diverse dexterous manipulation tasks.

Robotic control policies learned from human demonstrations have achieved impressive results in many real-world applications. However, in scenarios where initial performance is not satisfactory, as is often the case in novel open-world settings, such behavioral cloning (BC)-learned policies typically require collecting additional human demonstrations to further improve their behavior---an expensive and time-consuming process. In contrast, reinforcement learning (RL) holds the promise of enabling autonomous online policy improvement, but often falls short of achieving this due to the large number of samples it typically requires. In this work we take steps towards enabling fast autonomous adaptation of BC-trained policies via efficient real-world RL. Focusing in particular on diffusion policies---a state-of-the-art BC methodology---we propose *diffusion steering via reinforcement learning* (DSRL): adapting the BC policy by running RL over its latent-noise space. We show that DSRL is highly sample efficient, requires only black-box access to the BC policy, and enables effective real-world autonomous policy improvement. Furthermore, DSRL avoids many of the challenges associated with finetuning diffusion policies, obviating the need to modify the weights of the base policy at all. We demonstrate DSRL on simulated benchmarks, real-world robotic tasks, and for adapting pretrained generalist policies, illustrating its sample efficiency and effective performance at real-world policy improvement.

Multi-agent reinforcement learning (MARL) has emerged as a promising solution for learning complex and scalable coordination behaviors in multi-robot systems. However, established MARL platforms (e.g., SMAC and MPE) lack robotics relevance and hardware deployment, leaving multi-robot learning researchers to develop bespoke environments and hardware testbeds dedicated to the development and evaluation of their individual contributions. The Multi-Agent RL Benchmark and Learning Environment for the Robotarium (MARBLER) is an exciting recent step in providing a standardized robotics-relevant platform for MARL, by bridging the Robotarium testbed with existing MARL software infrastructure. However, MARBLER lacks support for parallelization and GPU/TPU execution, making the platform prohibitively slow compared to modern MARL environments and hindering adoption. We contribute JaxRobotarium, a Jax-powered end-to-end simulation, learning, deployment, and benchmarking platform for the Robotarium. JaxRobotarium enables rapid training and deployment of multi-robot reinforcement learning (MRRL) policies with realistic robot dynamics and safety constraints, supporting both parallelization and hardware acceleration. Our generalizable learning interface provides an easy-to-use integration with SOTA MARL libraries (e.g., JaxMARL). In addition, JaxRobotarium includes eight standardized coordination scenarios, including four novel scenarios that bring established MARL benchmark tasks (e.g., RWARE and Level-Based Foraging) to a realistic robotics setting. We demonstrate that JaxRobotarium retains high simulation fidelity while achieving dramatic speedups over baseline (20x in training and 150x in simulation), and provides an open-access sim-to-real evaluation pipeline through the Robotarium testbed, accelerating and democratizing access to multi-robot learning research and evaluation.

Scalable and reproducible policy evaluation has been a long-standing challenge in robot learning: evaluations are critical to assess progress and build better policies, but evaluation in the real world, especially at a scale that would provide statistically reliable results, is costly in terms of human time and hard to obtain. Evaluation of increasingly generalist robot policies requires an increasingly diverse repertoire of evaluation environments, making the evaluation bottleneck even more pronounced. To make real-world evaluation of robotic policies more practical, we propose AutoEval, a system to autonomously evaluate generalist robot policies around the clock with minimal human intervention. Users interact with AutoEval by submitting evaluation jobs to the AutoEval queue, much like how software jobs are submitted with a cluster scheduling system, and AutoEval will schedule the policies for evaluation within a framework supplying automatic success detection and automatic scene resets. We show that AutoEval can nearly fully eliminate human involvement in the evaluation process, permitting around the clock evaluations, and the evaluation results correspond closely to ground truth evaluations conducted by hand. To facilitate the evaluation of generalist policies in the robotics community, we provide public access to multiple AutoEval scenes in the popular BridgeData robot setup with WidowX robot arms. In the future, we hope that AutoEval scenes can be set up across institutions to form a diverse and distributed evaluation network.

Vision-Language-Action (VLA) models have become a cornerstone in robotic policy learning, leveraging large-scale multimodal data for robust and scalable control. However, existing VLA frameworks primarily address short-horizon tasks, and their effectiveness on long-horizon, multi-step robotic manipulation remains limited due to challenges in skill chaining and subtask dependencies. In this work, we introduce Long-VLA, the first end-to-end VLA model specifically designed for long-horizon robotic tasks. Our approach features a novel phase-aware input masking strategy that adaptively segments each subtask into moving and interaction phases, enabling the model to focus on phase-relevant sensory cues and enhancing subtask compatibility. This unified strategy preserves the scalability and data efficiency of VLA training, and our architecture-agnostic module can be seamlessly integrated into existing VLA models. We further propose the L-CALVIN benchmark to systematically evaluate long-horizon manipulation. Extensive experiments on both simulated and real-world tasks demonstrate that Long-VLA significantly outperforms prior state-of-the-art methods, establishing a new baseline for long-horizon robotic control.

Solving complex long-horizon robotic manipulation problems requires sophisticated high-level planning capabilities, the ability to reason about the physical world, and reactively choose appropriate motor skills. Vision-language models (VLMs) pretrained on Internet data could in principle offer a framework for tackling such problems. However, in their current form, VLMs lack both the nuanced understanding of intricate physics required for robotic manipulation and the ability to reason over long horizons to address error compounding issues. In this paper, we introduce a novel test-time computation framework that enhances VLMs' physical reasoning capabilities for multi-stage manipulation tasks. At its core, our approach iteratively improves a pretrained VLM with a ``reflection'' mechanism - it uses a generative model to imagine future world states, leverages these predictions to guide action selection, and critically reflects on potential suboptimalities to refine its reasoning. Experimental results demonstrate that our method significantly outperforms several state-of-the-art commercial VLMs as well as other post-training approaches such as Monte Carlo Tree Search (MCTS). Videos are available at https://corl2025-reflectvlm.github.io.

We introduce a novel History-Aware VErifier (HAVE) to disambiguate uncertain scenarios online by leveraging past interactions. Robots frequently encounter visually ambiguous objects whose manipulation outcomes remain uncertain until physically interacted with. While generative models alone could theoretically adapt to such ambiguity, in practice they obtain suboptimal performance in ambiguous cases, even when conditioned on action history. To address this, we propose explicitly decoupling action generation from verification: we use an unconditional diffusion-based generator to propose multiple candidate actions and employ our history-aware verifier to select the most promising action by reasoning about past interactions. Through theoretical analysis, we demonstrate that employing a verifier significantly improves expected action quality. Empirical evaluations and analysis across multiple simulated and real-world environments including articulated objects, multi-modal doors, and uneven object pick-up confirm the effectiveness of our method and improvements over baselines.

To perform autonomous visual search for environmental monitoring, a robot may leverage satellite imagery as a prior map. This can help inform coarse, high level search and exploration strategies, even when such images lack sufficient resolution to allow fine-grained, explicit visual recognition of targets. However, there are some challenges to overcome with using satellite images to direct visual search. For one, targets that are unseen in satellite images are underrepresented (compared to real life) in most existing datasets, and thus vision models trained on these datasets fail to reason effectively based on indirect visual cues. Furthermore, approaches which leverage large Vision Language Models (VLMs) for generalization may yield inaccurate outputs due to hallucination, leading to inefficient search. To address these challenges, we introduce Search-TTA, a multimodal test-time adaptation framework that can accept text and/or image input. First, we pretrain a remote sensing image encoder to align with CLIP’s visual encoder to output probability distributions of target presence used for visual search. Second, our framework dynamically refines CLIP’s predictions during search using a test-time adaptation mechanism. Through a feedback loop inspired by Spatial Poisson Point Processes, gradient updates (weighted by uncertainty) are used to correct (potentially inaccurate) predictions and improve search performance. To validate Search-TTA’s performance, we curate a visual search dataset based on internet-scale ecological data. We find that Search-TTA improves planner performance by up to 9.7%, particularly in cases with poor initial CLIP predictions. It also achieves comparable performance to state-of-the-art VLMs. Finally, we deploy Search-TTA on a real UAV via hardware-in-the-loop testing, by simulating its operation within a large-scale simulation that provides onboard sensing.

Aligning robot behavior with human preferences is crucial for deploying embodied AI agents in human-centered environments. A promising solution is interactive imitation learning from human intervention, where a human expert observes the policy's execution and provides interventions as feedback. However, existing methods often fail to utilize the prior policy efficiently to facilitate learning, thus hindering sample efficiency. In this work, we introduce Maximum-Entropy Residual-Q Inverse Reinforcement Learning, designed for sample-efficient alignment from human intervention. Instead of inferring the complete human behavior characteristics, MEReQ infers a residual reward function that captures the discrepancy between the human expert's and the prior policy's underlying reward functions. It then employs Residual Q-Learning (RQL) to align the policy with human preferences using this residual reward function. Extensive evaluations on simulated and real-world tasks demonstrate that MEReQ achieves sample-efficient policy alignment from human intervention compared to other baselines.

Teleoperating humanoid robots in a whole-body manner marks a fundamental step toward developing general-purpose robotic intelligence, with human motion providing an ideal interface for controlling all degrees of freedom. Yet, most current humanoid teleoperation systems fall short of enabling coordinated whole-body behavior, typically limiting themselves to isolated locomotion or manipulation tasks. We present the Teleoperated Whole-Body Imitation System (TWIST), a system for humanoid teleoperation through whole-body motion imitation. We first generate reference motion clips by retargeting human motion capture data to the humanoid robot. We then develop a robust, adaptive, and responsive whole-body controller using a combination of reinforcement learning and behavior cloning (RL+BC). Through systematic analysis, we demonstrate how incorporating privileged future motion frames and real-world motion capture (MoCap) data improves tracking accuracy. TWIST enables real-world humanoid robots to achieve unprecedented, versatile, and coordinated whole-body motor skills—spanning whole-body manipulation, legged manipulation, locomotion, and expressive movement—using a single unified neural network controller.

Foundation models can provide robust high-level reasoning on appropriate safety interventions in hazardous scenarios beyond a robot's training data, i.e. out-of-distribution (OOD) failures. However, due to the high inference latency of Large Vision and Language Models, current methods rely on manually defined intervention policies to enact fallbacks, thereby lacking the ability to plan generalizable, semantically safe motions. To overcome these challenges we present FORTRESS, a framework that generates and reasons about semantically safe fallback strategies in real time to prevent OOD failures. At a low frequency in nominal operations, FORTRESS uses multi-modal reasoners to identify goals and anticipate failure modes. When a runtime monitor triggers a fallback response, FORTRESS rapidly synthesizes plans to fallback goals while inferring and avoiding semantically unsafe regions in real time. By bridging open-world, multi-modal reasoning with dynamics-aware planning, we eliminate the need for hard-coded fallbacks and human safety interventions. FORTRESS outperforms on-the-fly prompting of slow reasoning models in safety classification accuracy on synthetic benchmarks and real-world ANYmal robot data, and further improves system safety and planning success in simulation and on quadrotor hardware for urban navigation. Website and code can be found at https://submfort.github.io/fortress/.

In this work, we propose a novel flow field-based motion planning method that drives a robot from any initial state to a desired reference trajectory such that it converges to the trajectory's end point. Despite demonstrated efficacy in using Koopman operator theory for modeling dynamical systems, Koopman does not inherently enforce convergence to desired trajectories nor to specified goals, a requirement when learning from demonstrations (LfD). We present KoopMotion which represents motion flow fields as dynamical systems, parameterized by Koopman Operators, and leverages the divergence properties of the learnt flow fields to obtain smooth motion fields that converge to a desired reference trajectory when the robot is placed away from the desired trajectory, and tracks the trajectory until the end point. To demonstrate the effectiveness of our approach, we show evaluations of KoopMotion on the LASA human handwriting dataset, including spectral analysis. We also perform experiments on a physical robot, verifying KoopMotion on a miniature autonomous surface vehicle operating in a non-static fluid flow environment. Our approach is highly sample efficient in both space and time, requiring only 3\% of the LASA dataset to generate dense motion plans. Additionally, KoopMotion provides a significant improvement over baselines when comparing metrics that measure spatial and temporal dynamics modeling efficacy.

Pointing gestures provide a natural and efficient way to communicate spatial information in human-machine interaction, yet their potential for 3D object referral remains largely under-explored. To fill this gap, we introduce the task of pointing-based 3D segmentation. In this task, given an image of a person pointing at an object and the 3D point cloud of the environment, the goal is to predict the 3D segmentation mask of the referred object. To enable the standardized evaluation of this task, we introduce POINTR3D, a curated dataset of over 65,000 frames captured with three cameras across four indoor scenes, featuring diverse pointing scenarios. Each frame is annotated with the information of the active hand, the corresponding object ID, and the 3D segmentation mask of the object. To showcase the application of the proposed dataset, we further introduce Pointing3D, a transformer-based architecture that predicts the pointing direction from RGB images and uses this prediction as a prompt to segment the referred object in the 3D point cloud. Experimental results show that Pointing3D outperforms other strong baselines we introduce and lays the groundwork for future research. The dataset, source code, and evaluation tools will be made publicly available to support further research in this area, enabling a natural human-machine interaction.

Hand–object motion-capture (MoCap) repositories provide abundant, contact-rich human demonstrations for scaling dexterous manipulation on robots. Yet demonstration inaccuracy and embodiment gaps between human and robot hands challenge direct policy learning. Existing pipelines adapt a three-stage workflow: retargeting, tracking, and residual correction. This multi-step process may not fully utilize demonstrations and can introduce compound errors. We introduce Reference-Scoped Exploration (RSE), a unified, single-loop optimization that integrates retargeting and tracking to train a scalable robot control policy directly from MoCap. Instead of treating demonstrations as strict ground truth, we view them as soft guidance. From raw demonstrations, we construct adaptive spatial scopes—time-varying termination boundaries, and reinforcement learning promotes the policy to stay within these envelopes while minimizing control effort. This holistic approach preserves demonstration intent, lets robot-specific strategies emerge, boosts robustness to noise, and scales effortlessly with large-scale demonstrations. We distill the scaled tracking policy into a vision-based, skill-conditioned generative control policy. This distilled policy captures diverse manipulation skills within a rich latent representation, enabling generalization across various objects and real-world robotic manipulation.

Articulated object recognition -- the task of identifying both the geometry and kinematic joints of objects with movable parts -- is essential for enabling robots to interact with everyday objects such as doors and laptops. However, existing approaches often rely on strong assumptions, such as a known number of articulated parts; require additional inputs, such as depth images; or involve complex intermediate steps that can introduce potential errors -- limiting their practicality in real-world settings. In this paper, we introduce **ScrewSplat**, a simple end-to-end method that operates solely on RGB observations. Our approach begins by randomly initializing screw axes, which are then iteratively optimized to recover the object’s underlying kinematic structure. By integrating with Gaussian Splatting, we simultaneously reconstruct the 3D geometry and segment the object into rigid, movable parts. We demonstrate that our method achieves state-of-the-art recognition accuracy across a diverse set of articulated objects, and further enables zero-shot, text-guided manipulation using the recovered kinematic model.

Imitation-based policy training for long-horizon manipulation tasks involving multi-step object interactions is often susceptible to compounding action errors. Contemporary methods discover semantic subgoals embedded within the overall task, decomposing the overall task into tractable shorter-horizon goal-conditioned policy learning. However, policy deployment requires iteratively estimating $\textit{which}$ subgoal is being pursued and $\textit{when}$ it is achieved. We observe the brittleness of conventional $\textit{heuristic}$-based approaches (ad hoc threshold based), particularly for long-horizon imitation, since pursuing an incorrect subgoal can lead the robot policy to experience out of distribution states. In this work, we introduce two policy architectures for modeling subgoal transitions within a policy learning loop for long-horizon tasks. The first model autoregressively predicts the likelihood of the next subgoal transition, while the second uses cross-attention (via a transformer-based architecture) and implicitly models smooth and continuous transitions. We evaluate our models on $25$ simulated tasks on Franka Kitchen, $6$ real-world table-top tasks and $18$ simulated tasks on a new corpus (Franka-Long Horizon Tasks (LHT)) focused on tasks with rich object interactions over long episode lengths. Experimental results show significant improvements in learning efficacy, task success rates and generalization to out-of-distribution settings- extending horizon lengths for imitating manipulation tasks $\textit{from long to long(er)}$.

Ensuring the robustness of robotic systems is crucial for their deployment in safety-critical domains. Failure discovery, or falsification, is a widely used approach for evaluating robustness, with recent advancements focusing on improving sample efficiency and generalization through probabilistic sampling techniques and learning-theoretic approaches. However, existing methods typically rely on explicitly defined analytical cost functions to characterize failures, often overlooking the underlying causes and diversity of discovered failure scenarios. In this work, we propose a novel failure discovery framework that integrates contextual reasoning in the falsification process, specifically tailored for high evaluation-cost applications. Our method incorporates expert-in-the-loop feedback to construct a probabilistic surrogate model of failures using Bayesian inference. This model is iteratively refined and leveraged to guide an active learning strategy that prioritizes the discovery of diverse failure cases. We empirically validate our approach across a range of tasks for high-cost contextual falsification in robotic manipulation and autonomous driving.

Generative models based on flow matching offer significant potential for learning robot policies, particularly in generating high-dimensional, dexterous behaviors that are conditioned on diverse observations. In this work, we introduce ManiFlow, an advanced flow matching model specifically designed to support dexterous manipulation tasks. ManiFlow improves over flow matching both in the learning procedure and in the model architecture, resulting in better robustness and efficacy. It consistently exhibits strong generalization capabilities, outperforming existing state-of-the-art robot learning methods on a wide range of benchmarks. We also demonstrate the powerful capabilities of ManiFlow in solving complex bimanual dexterous manipulation challenges.