CoRL 2025 论文列表

Abstract: Shared autonomy is an enabling technology that provides users with control authority over robots that would otherwise be difficult if not impossible to directly control. Yet, standard methods make assumptions that limit their adoption in practice—for example, prior knowledge of the user’s goals or the objective (i.e., reward) function that they wish to optimize, knowledge of the user’s policy, or query-level access to the user during training. Diffusion-based approaches to shared autonomy do not make such assumptions and instead only require access to demonstrations of desired behaviors, while allowing the user to maintain control authority. However, these advantages have come at the expense of high computational complexity, which has made real-time shared autonomy all but impossible. To overcome this limitation, we propose Consistency Shared Autonomy (CSA), a shared autonomy framework that employs a consistency model-based formulation of diffusion. Key to CSA is that it employs the distilled probability flow of ordinary differential equations (PF ODE) to generate high-fidelity samples in a single step. This results in inference speeds significantly than what is possible with previous diffusion-based approaches to shared autonomy, enabling real-time assistance in complex domains with only a single function evaluation. Further, by intervening on flawed actions at intermediate states of the PF ODE, CSA enables varying levels of assistance. We evaluate CSA on a variety of challenging simulated and real-world robot control problems, demonstrating significant improvements over state-of-the-art methods both in terms of task performance and computational efficiency.

Robust robot manipulation in unstructured environments often requires understanding object properties that extend beyond geometry, such as material or compliance—properties that can be challenging to infer using vision alone. Multimodal haptic sensing provides a promising avenue for inferring such properties, yet progress has been constrained by the lack of large, diverse, and realistic haptic datasets. In this work, we introduce the CLAMP device, a low-cost (< $200) sensorized reacher-grabber designed to collect large-scale, in-the-wild multimodal haptic data from non-expert users in everyday settings. We deployed 16 CLAMP devices to 41 participants, resulting in the CLAMP dataset, the largest open-source multimodal haptic dataset to date, comprising 12.3 million datapoints across 5357 household objects. Using this dataset, we train a haptic encoder that can infer material and compliance object properties from multimodal haptic data. We leverage this encoder to create the CLAMP model, a visuo-haptic perception model for material recognition that generalizes to novel objects and three robot embodiments with minimal finetuning. We also demonstrate the effectiveness of our model in three real-world robot manipulation tasks: sorting recyclable and non-recyclable waste, retrieving objects from a cluttered bag, and distinguishing overripe from ripe bananas. Our results show that large-scale, in-the-wild haptic data collection can unlock new capabilities for generalizable robot manipulation.

Incorporating task-specific priors within a policy or network architecture is crucial for enhancing safety and improving representation and generalization in robotic control problems. Differentiable Model Predictive Control (MPC) layers have proven effective for embedding these priors, such as constraints and cost functions, directly within the architecture, enabling end-to-end training. However, current methods often treat the solver and the neural network as separate, independent entities, leading to suboptimal integration. In this work, we propose a novel approach that co-develops the solver and architecture unifying the optimization solver and network inference problems. Specifically, we formulate this as a \textit{joint fixed-point problem} over the coupled network outputs and necessary conditions of the optimization problem. We solve this problem in an iterative manner where we alternate between network forward passes and optimization iterations. Through extensive ablations in various robotic control tasks, we demonstrate that our approach results in richer representations and more stable training, while naturally accommodating warm starting, a key requirement for MPC.

Vision is well-known for its use in manipulation, especially using visual servoing. To make it robust, multiple cameras are needed to expand the field of view. That is computationally challenging. Merging multiple views and using Q-learning allows the design of more effective representations and optimization of sample efficiency. Such a solution might be expensive to deploy. To mitigate this, we introduce a merge and disentanglement (MAD) algorithm that efficiently merges views to increase sample efficiency while augmenting with single-view features to allow lightweight deployment and ensure robust policies. We demonstrate the efficiency and robustness of our approach using Meta-World and ManiSkill3.

Embodied foundation models are gaining increasing attention for their zero-shot generalization, scalability, and adaptability to new tasks through few-shot post-training. However, existing models rely heavily on real-world data, which is costly and labor-intensive to collect. Synthetic data offers a cost-effective alternative, yet its potential remains largely underexplored. To bridge this gap, we explore the feasibility of training Vision-Language-Action (VLA) models entirely with large-scale synthetic action data. We curate SynGrasp-1B, a billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and extensive domain randomization. Building on this, we present GraspVLA, a VLA model pretrained on large-scale synthetic action data as a foundational model for grasping tasks. GraspVLA integrates autoregressive perception tasks and a flow-matching-based action generation into a unified Chain-of-Thought process, enabling joint training on synthetic action data and Internet semantics data. This design helps mitigate sim-to-real gaps and facilitates the transfer of learned actions to a broader range of Internet-covered objects, achieving open-vocabulary generalization in grasping. Extensive evaluations across real-world and simulation benchmarks demonstrate GraspVLA’s advanced zero-shot generalizability and few-shot adaptability to specific human preferences. We will release SynGrasp-1B dataset and pre-trained weights to benefit the community.

Generalizing locomotion policies across diverse legged robots with varying morphologies is a key challenge due to differences in observation/action dimensions and system dynamics. In this work, we propose \textit{Multi-Loco}, a novel unified framework combining a morphology-agnostic generative diffusion model with a lightweight residual policy optimized via reinforcement learning (RL). The diffusion model captures morphology-invariant locomotion patterns from diverse cross-embodiment datasets, improving generalization and robustness. The residual policy is shared across all embodiments and refines the actions generated by the diffusion model, enhancing task-aware performance and robustness for real-world deployment. We evaluated our method with a rich library of four legged robots in both simulation and real-world experiments. Compared to a standard RL framework with PPO, our approach - replacing the Gaussian policy with a diffusion model and residual term - achieves a 10.35\% average return improvement, with gains up to 13.57\% in wheeled-biped locomotion tasks. These results highlight the benefits of cross-embodiment data and composite generative architectures in learning robust, generalized locomotion skills.

In unstructured environments, robotic manipulation tasks involving objects with constrained motion trajectories—such as door opening—often experience discrepancies between the robot's vision-guided end-effector trajectory and the object's constrained motion path. Such discrepancies generate unintended harmful forces, which, if exacerbated, may lead to task failure and potential damage to the manipulated objects or the robot itself. To address this issue, this paper introduces a novel diffusion framework, termed SafeDiff. Unlike conventional methods that sequentially fuse visual and tactile data to predict future robot states, our approach generates a prospective state sequence based on the current robot state and visual context observations, using real-time force feedback as a calibration signal. This implicitly adjusts the robot’s state within the state space, enhancing operational success rates and significantly reducing harmful forces during manipulation, thus ensuring manipulation force safety. Additionally, we develop a large-scale simulation dataset named SafeDoorManip50k, offering extensive multimodal data to train and evaluate the proposed method. Extensive experiments show that our visual-tactile model substantially mitigates the risk of harmful forces in the door opening task, across both simulated and real-world settings.

Long-horizon contact-rich tasks are challenging to learn with reinforcement learning, due to ineffective exploration of high-dimensional state spaces with sparse rewards. The learning process often gets stuck in local optimum and demands task-specific reward fine-tuning for complex scenarios. In this work, we propose a structured framework that leverages privileged actions with curriculum learning, enabling the policy to efficiently acquire long-horizon skills without relying on extensive reward engineering or reference trajectories. Specifically, we use privileged actions in simulation with a general training procedure that would be infeasible to implement in real-world scenarios. These privileges include relaxed constraints and virtual forces that enhance interaction and exploration with objects. Our results successfully achieve complex multi-stage long-horizon tasks that naturally combine non-prehensile manipulation with grasping to lift objects from non-graspable poses. We demonstrate generality by maintaining a parsimonious reward structure and showing convergence to diverse and robust behaviors across various environments. Our approach outperforms state-of-the-art methods in these tasks, converging to solutions where others fail.

Robotic agents operating in dynamic, partially observable environments often benefit from teammate assistance. We address the challenge of determining when and how to assist in multi-robot systems where agents can modify the physical environment, such as moving obstacles that block perception or manipulation. For robots using online POMDP planning, evaluating assistance impacts requires computationally intensive policy evaluation, making real-time decisions difficult. We formulate Value of Assistance (VOA) for POMDP agents and develop efficient heuristics that approximate VOA without requiring complete policy evaluation. Our empirical evaluation on both a standard POMDP benchmark and a collaborative manipulation task demonstrates that our Full Information heuristic enables real-time assistance decisions while maintaining sufficient accuracy for effective helping action selection.

This paper addresses the challenge of occluded robot grasping, i.e. grasping in situations where the desired grasp poses are kinematically infeasible due to environmental constraints such as surface collisions. Existing RL methods struggle with task complexity, and collecting expert demonstrations is often impractical. Instead, inspired by human bimanual manipulation strategies, where two hands coordinate to stabilise and reorient objects, we focus on a bimanual robotic setup to tackle this challenge. In particular, we introduce Constraint-based Manipulation for Bimanual Occluded Grasping (COMBO-Grasp), an approach which leverages two coordinated policies: a constraint policy trained using self-supervised datasets to generate stabilising poses and a grasping policy trained using RL that reorients and grasps the target object. A key contribution lies in value function-guided policy coordination, where gradients from a jointly trained value function refine the constraint policy during RL training to improve bimanual coordination and task performance. Lastly, COMBO-Grasp employs teacher-student policy distillation to effectively deploy vision-based policies in real-world environments. Experiments show that COMBO-Grasp significantly outperforms baselines and generalises to unseen objects in both simulation and real environments.

Robot navigation in dynamic, human-centered environments requires socially-compliant decisions grounded in robust scene understanding, including spatiotemporal awareness and the ability to interpret human intentions. Recent Vision-Language Models (VLMs) show exhibit promising capabilities such as object recognition, common-sense reasoning, and contextual understanding—that align with the nuanced requirements of social robot navigation. However, it remains unclear whether VLMs can reliably perform the complex spatiotemporal reasoning and intent inference needed for safe and socially compliant robot navigation. In this paper, we introduce the Social Navigation Scene Understanding Benchmark (SocialNav-SUB), a Visual Question Answering (VQA) dataset and benchmark designed to evaluate VLMs for scene understanding in real-world social robot navigation scenarios. SocialNav-SUB provides a unified framework for evaluating VLMs against human and rule-based baselines across VQA tasks requiring spatial, spatiotemporal, and social reasoning in social robot navigation. Through experiments with state-of-the-art VLMs, we find that while the best-performing VLM achieves an encouraging probability of agreeing with human answers, it still underperforms a simpler rule-based approach and human consensus, indicating critical gaps in social scene understanding of current VLMs. Our benchmark sets the stage for further research on foundation models for social robot navigation, offering a framework to explore how VLMs can be tailored to meet real-world social robot navigation needs. We will open source the code and release the benchmark.

Adaptive teaming—the capability of agents to effectively collaborate with unfamiliar teammates without prior coordination—is widely explored in virtual video games but overlooked in real-world multi-robot contexts. Yet, such adaptive collaboration is crucial for real-world applications, including border surveillance, search-and-rescue, and counter-terrorism operations. To address this gap, we introduce AT-Drone, the first dedicated benchmark explicitly designed to facilitate comprehensive training and evaluation of adaptive teaming strategies in multi-drone pursuit scenarios. AT-Drone makes the following key contributions: (1) An adaptable simulation environment configurator that enables intuitive and rapid setup of adaptive teaming multi-drone pursuit tasks, including four predefined pursuit environments. (2) A streamlined real-world deployment pipeline that seamlessly translates simulation insights into practical drone evaluations using edge devices (such as Jetson Orin Nano) and Crazyflie drones. (3) A novel algorithm zoo integrated with a distributed training framework, featuring diverse algorithms explicitly tailored, for the first time, to multi-pursuer and multi-evader drone pursuit task. (4) Standardized evaluation protocols with newly designed unseen drone zoos, explicitly designed to rigorously assess the performance of adaptive teaming. Comprehensive experimental evaluations across four progressively challenging multi-drone pursuit scenarios confirm AT-Drone's effectiveness in advancing adaptive teaming research. Real-world drone experiments further validate its practical feasibility and utility for realistic robotic operations. Videos, code and weights are available at \url{https://sites.google.com/view/at-drone}.

While large-scale robot datasets have propelled recent progress in imitation learning, learning from smaller task specific datasets remains critical for deployment in new environments and unseen tasks. One such approach to few-shot imitation learning is retrieval-based imitation learning, which extracts relevant samples from large, widely available prior datasets to augment a limited demonstration dataset. To determine the relevant data from prior datasets, retrieval-based approaches most commonly calculate a prior data point's minimum distance to a point in the target dataset in latent space. While retrieval-based methods have shown success using this metric for data selection, we demonstrate its equivalence to the limit of a Gaussian kernel density (KDE) estimate of the target data distribution. This reveals two shortcomings of the retrieval rule used in prior work. First, it relies on high-variance nearest neighbor estimates that are susceptible to noise. Second, it does not account for the distribution of prior data when retrieving data. To address these issues, we introduce Importance Weighted Retrieval (IWR), which estimates importance weights, or the ratio between the target and prior data distributions for retrieval, using Gaussian KDEs. By considering the probability ratio, IWR overcomes the bias of previous selection rules, and by using reasonable modeling parameters, IWR effectively smooths estimates using all data points. Across both simulation environments and real-world evaluations on the Bridge dataset we find that our method, IWR, consistently improves performance of existing retrieval-based methods, despite only requiring minor modifications.

Operating home appliances, among the most common tools in every household, is a critical capability for assistive home robots. This paper presents ApBot, a robot system that operates novel household appliances by “reading” their user manuals. ApBot faces multiple challenges: (i) infer goal-conditioned partial policies from their unstructured, textual descriptions in a user manual document, (ii) ground the policies to the appliance in the physical world, and (iii) execute the policies reliably over potentially many steps, despite compounding errors. To tackle these challenges, ApBot constructs a structured, symbolic model of an appliance from its manual, with the help of a large vision-language model (VLM). It grounds the symbolic actions visually to control panel elements. Finally, ApBot closes the loop by updating the model based on visual feedback. Our experiments show that across a wide range of simulated and real-world appliances, ApBot achieves consistent and statistically significant improvements in task success rate, compared with state-of-the-art large VLMs used directly as control policies. These results suggest that a structured internal representations plays an important role in robust robot operation of home appliances, especially, complex ones.

Learning robot policies that capture multimodality in the training data has been a long-standing open challenge for behavior cloning. Recent approaches tackle the problem by modeling the conditional action distribution with generative models. One of these approaches is Diffusion Policy, which relies on a diffusion model to denoise random points into robot action trajectories. While achieving state-of-the-art performance, it has two main drawbacks that may lead the robot out of the data distribution during policy execution. First, the stochasticity of the denoising process can highly impact on the quality of generated trajectory of actions. Second, being a supervised learning approach, it can learn data outliers from the dataset used for training. Recent work focuses on mitigating these limitations by combining Diffusion Policy either with large-scale training or with classical behavior cloning algorithms. Instead, we propose KDPE, a Kernel Density Estimation-based strategy that filters out potentially harmful trajectories output of Diffusion Policy while keeping a low test-time computational overhead. For Kernel Density Estimation, we propose a manifold-aware kernel to model a probability density function for actions composed of end-effector Cartesian position, orientation, and gripper state. KDPE overall achieves better performance than Diffusion Policy on simulated single-arm RoboMimic and MimicGen tasks, and on three real robot experiments:PickPlush, a tabletop grasping task, CubeSort, a multimodal pick and place task, and CoffeeMaking, a task that requires long-horizon capabilities and precise execution. The code will be released upon acceptance and additional material is provided on our anonymized project page:https://kdpe-robotics.github.io.

In order for robots to be useful, they must perform practically relevant tasks in the real world, outside of the lab. While vision-language-action (VLA) models have demonstrated impressive results for end-to-end robot control, it remains an open question how far such models can generalize in the wild. We describe $\pi_{0.5}$, a new model based on $\pi_0$ that uses co-training on heterogeneous tasks to enable broad generalization. $\pi_{0.5}$ uses data from multiple robots, high-level semantic prediction, web data, and other sources to enable broadly generalizable real-world robotic manipulation. Our system uses a combination of co-training and hybrid multi-modal examples that combine image observations, language commands, object detections, semantic subtask prediction, and low-level actions. Our experiments show that this kind of knowledge transfer is essential for effective generalization, and we demonstrate for the first time that an end-to-end learning-enabled robotic system can perform long-horizon and dexterous manipulation skills, such as cleaning a kitchen or bedroom, in entirely new homes.

Reliable localization is critical for robot navigation, yet many existing systems assume that all viewpoints along a trajectory are equally informative. In practice, localization becomes unreliable when the robot observes unmapped, ambiguous, or uninformative regions. To address this, we present ActLoc, an active viewpoint-aware planning framework for enhancing localization accuracy for general robot navigation tasks. At the core of ActLoc is an attention-based model trained at scale for viewpoint selection. This model encodes a metric map of the scene, along with camera poses used during map construction, and estimates localization accuracy over camera pitch and yaw directions at arbitrary 3D waypoint in space. This per-point accuracy distribution is integrated into the path planning process, allowing the robot to actively choose camera orientation that maximize localization robustness while respecting task and motion constraints. ActLoc achieves state-of-the-art performance in single-viewpoint selection task, and generalizes effectively to full-trajectory planning. It provides a modular enhancement to a wide range of navigation and inspection tasks in structured environments.

Real-world household tasks present significant challenges for mobile manipulation robots. An analysis of existing robotics benchmarks reveals that successful task performance hinges on three key whole-body control capabilities: bimanual coordination, stable and precise navigation, and extensive end-effector reachability. Achieving these capabilities requires careful hardware design, but the resulting system complexity further complicates visuomotor policy learning. To address these challenges, we introduce the BEHAVIOR Robot Suite (BRS), a comprehensive framework for whole-body manipulation in diverse household tasks. Built on a bimanual, wheeled robot with a 4-DoF torso, BRS integrates a cost-effective whole-body teleoperation interface for data collection and a novel algorithm for learning whole-body visuomotor policies. We evaluate BRS on five challenging household tasks that not only emphasize the three core capabilities but also introduce additional complexities, such as long-range navigation, interaction with articulated and deformable objects, and manipulation in confined spaces. We believe that BRS's integrated robotic embodiment, data collection interface, and learning framework mark a significant step toward enabling real-world whole-body manipulation for everyday household tasks. BRS is open-sourced at https://behavior-robot-suite.github.io/.

Standard evaluation protocols in robotic manipulation typically assess policy performance over curated, in-distribution test sets, offering limited insight into how systems fail under plausible variation. We introduce a red-teaming framework that probes robustness through object-centric geometric perturbations, automatically generating CrashShapes---structurally valid, user-constrained mesh deformations that trigger catastrophic failures in pre-trained manipulation policies. The method integrates a Jacobian field–based deformation model with a gradient-free, simulator-in-the-loop optimization strategy. Across insertion, articulation, and grasping tasks, our approach consistently discovers deformations that collapse policy performance, revealing brittle failure modes missed by static benchmarks. By combining task-level policy rollouts with constraint-aware shape exploration, we aim to build a general purpose framework for structured, object-centric robustness evaluation in robotic manipulation. We additionally show that fine-tuning on individual CrashShapes, a process we refer to as blue-teaming, improves task success by up to 60 percentage points on those shapes, while preserving performance on the original object, demonstrating the utility of red-teamed geometries for targeted policy refinement. Finally, we validate both red-teaming and blue-teaming results with a real robotic arm, observing that simulated CrashShapes reduce task success from 90\% to as low as 22.5\%, and that blue-teaming recovers performance to up to 90\% on the corresponding real-world geometry---closely matching simulation outcomes.