CoRL 2025 论文列表

As robots become increasingly capable of operating over extended periods—spanning days, weeks, and even months—they are expected to accumulate knowledge of their environments and leverage this experience to assist humans more effectively. This paper studies the problem of Long-term Active Embodied Question Answering (LA-EQA), a new task in which a robot must both recall past experiences and actively explore its environment to answer complex, temporally-grounded questions. Unlike traditional EQA settings, which typically focus either on understanding the present environment alone or on recalling a single past observation, LA-EQA challenges an agent to reason over past, present, and possible future states, deciding when to explore, when to consult its memory, and when to stop gathering observations and provide a final answer. Standard EQA approaches based on large models struggle in this setting due to limited context windows, absence of persistent memory, and an inability to combine memory recall with active exploration. To address this, we propose a structured memory system for robots, inspired by the mind palace method from cognitive science. Our method encodes episodic experiences as scene-graph-based world instances, forming a reasoning and planning algorithm that enables targeted memory retrieval and guided navigation. To balance the exploration-recall trade-off, we introduce value-of-information-based stopping criteria that determines when the agent has gathered sufficient information. We evaluate our method on real-world experiments and introduce a new benchmark that spans popular simulation environments and actual industrial sites. Our approach significantly outperforms state-of-the-art baselines, yielding substantial gains in both answer accuracy and exploration efficiency.

Dexterous grasping in cluttered scenes presents significant challenges due to diverse object geometries, occlusions, and potential collisions. Existing methods primarily focus on single-object grasping or grasp-pose prediction without interaction, which are insufficient for complex, cluttered scenes. Recent vision-language-action models offer a potential solution but require extensive real-world demonstrations, making them costly and difficult to scale. To address these limitations, we revisit the sim-to-real transfer pipeline and develop key techniques that enable zero-shot deployment in reality while maintaining robust generalization. We propose ClutterDexGrasp, a two-stage teacher-student framework for closed-loop target-oriented dexterous grasping in cluttered scenes. The framework features a teacher policy trained in simulation using clutter density curriculum learning, incorporating both a novel geometry- and spatially-embedded scene representation and a comprehensive safety curriculum, enabling general, dynamic, and safe grasping behaviors. Through imitation learning, we distill the teacher's knowledge into a student 3D diffusion policy (DP3) that operates on partial point cloud observations. To the best of our knowledge, this represents the first zero-shot sim-to-real closed-loop system for target oriented dexterous grasping in cluttered scenes, demonstrating robust performance across diverse objects and layouts.

Gaussian splatting has emerged as a powerful tool for high-fidelity reconstruction of dynamic scenes. However, existing methods primarily rely on implicit motion representations, such as encoding motions into neural networks or per-Gaussian parameters, which makes it difficult to further manipulate the reconstructed motions. This lack of explicit controllability limits existing methods to replaying recorded motions only, which hinders a wider application. To address this, we propose Motion Blender Gaussian Splatting (MB-GS), a novel framework that uses motion graph as an explicit and sparse motion representation. The motion of graph links is propagated to individual Gaussians via dual quaternion skinning, with learnable weight painting functions determining the influence of each link. The motion graphs and 3D Gaussians are jointly optimized from input videos via differentiable rendering. Experiments show that MB-GS achieves state-of-the-art performance on the iPhone dataset while being competitive on HyperNeRF. Additionally, we demonstrate the application potential of our method in animating novel object motions, synthesizing robot demonstrations through motion editing, and predicting robot actions through visual planning.

Robots often struggle to generalize from a single demonstration due to the lack of a transferable and interpretable spatial representation. In this work, we introduce TReF-6, a method that infers a simplified, abstracted 6DoF Task-Relevant Frame from a single trajectory. Our approach identifies an influence point purely from the trajectory geometry to define the origin for a local frame, which serves as a reference for parameterizing a Dynamic Movement Primitive (DMP). This influence point captures the task's spatial structure, extending the standard DMP formulation beyond start-goal imitation. The inferred frame is semantically grounded via a vision-language model and localized in novel scenes by Grounded-SAM, enabling functionally consistent skill generalization. We validate TReF-6 in simulation and demonstrate robustness to trajectory noise. We further deploy an end-to-end pipeline on real-world manipulation tasks, showing that TReF-6 supports one-shot imitation learning that preserves task intent across diverse object configurations.

We introduce Lucid-XR, a generative data engine for creating diverse and realistic-looking data to train real-world robot systems. At the core of Lucid-XR is vuer, a web-based physics simulation environment that runs directly on the XR headset, enabling internet-scale access to immersive, latency-free virtual interactions without requiring specialized equipment. The complete system integrates on-device physics simulation with on-device human-to-robot pose retargeting, that are further amplified by a physics-guided video generation pipeline commandable with natural language specifications. We demonstrate zero-shot sim-to-real transfer of robot visual policies, trained entirely on Lucid-XR's synthetic data, across bimanual and dexterous manipulation tasks that involve flexible materials, adhesive interaction between particles, and rigid body contact.

In this work, we unlock new capabilities in robot learning from neural trajectories, synthetic robot data generated from video world models. Our proposed recipe is simple, but powerful: we take the most recent state-of-the-art video generative models (world models), adapt them to the target robot embodiment, and generate new, synthetic robot data of the same task or even new behaviors. Since these video world models only generate videos, we explore two techniques of getting robot actions: extracting latent actions from a general-purpose latent action model and getting predicted actions from an inverse-dynamics model (IDM), giving flexibility across diverse scenarios. Our proposed approach unlocks behavior and environment generalization, allowing a humanoid robot to perform 20+ new behaviors in unseen environments while only collecting teleoperation data for pick and place in a single environment. By introducing a new world modeling benchmark, we demonstrate that stronger video world models directly correlate with improved downstream robot policy performance. This establishes a new scaling dimension beyond simply collecting additional teleoperation data, changing how we approach robot learning.

We focus on the task of identifying the location of target regions from a natural language instruction and a front camera image captured by a mobility. This task is challenging because it requires both existence prediction and segmentation mask generation, particularly for stuff-type target regions with ambiguous boundaries. Existing methods often underperform in handling stuff-type target regions, in addition to absent or multiple targets. To overcome these limitations, we propose GENNAV, which predicts target existence and generates segmentation masks for multiple stuff-type target regions. To evaluate GENNAV, we constructed a novel benchmark called GRiN-Drive, which includes three distinct types of samples: no-target, single-target, and multi-target. GENNAV achieved superior performance over baseline methods on standard evaluation metrics. Furthermore, we conducted real-world experiments with four automobiles operated in five geographically distinct urban areas to validate its zero-shot transfer performance. In these experiments, GENNAV outperformed baseline methods and demonstrated its robustness across diverse real-world environments.

Robot manipulation in unstructured environments requires efficient and reliable Swept Volume Collision Detection (SVCD) for safe motion planning. Traditional discrete methods potentially miss collisions between these points, whereas SVCD continuously checks for collisions along the entire trajectory. Existing SVCD methods typically face a trade-off between efficiency and accuracy, limiting practical use. In this paper, we introduce NeuralSVCD, a novel neural encoder-decoder architecture tailored to overcome this trade-off. Our approach leverages shape locality and temporal locality through distributed geometric representations and temporal optimization. This enhances computational efficiency without sacrificing accuracy. Comprehensive experiments show that NeuralSVCD consistently outperforms existing state-of-the-art SVCD methods in terms of both collision detection accuracy and computational efficiency, demonstrating its robust applicability across diverse robotic manipulation scenarios. Code and videos are available at https://neuralsvcd.github.io/.

There is growing interest in reinforcement learning (RL) methods that leverage the simulator's derivatives to improve learning efficiency. While early gradient-based approaches have demonstrated superior performance compared to derivative-free methods, accessing simulator gradients is often impractical due to their implementation cost or unavailability. Model-based RL (MBRL) can approximate these gradients via learned dynamics models, but the solver efficiency suffers from compounding prediction errors during training rollouts, which can degrade policy performance. We propose an approach that decouples trajectory generation from gradient computation: trajectories are unrolled using a simulator, while gradients are computed via backpropagation through a learned differentiable model of the simulator. This hybrid design enables efficient and consistent first-order policy optimization, even when simulator gradients are unavailable, as well as learning a critic from simulation rollouts, which is more accurate. Our method achieves the sample efficiency and speed of specialized optimizers such as SHAC, while maintaining the generality of standard approaches like PPO and avoiding ill behaviors observed in other first-order MBRL methods. We empirically validate our algorithm on benchmark control tasks and demonstrate its effectiveness on a real Go2 quadruped robot, across both quadrupedal and bipedal locomotion tasks.

With robots increasingly operating in human-centric environments, ensuring soft and safe physical interactions, whether with humans, surroundings, or other machines, is essential. While compliant hardware can facilitate such interactions, this work focuses on impedance controllers that allow torque-controlled robots to safely and passively respond to contact while accurately executing tasks. From inverse dynamics to quadratic programming based controllers, the effectiveness of these methods relies on accurate dynamics models of the robot and the object it manipulates. Any model mismatch results in task failures and unsafe behaviors. Thus, we introduce Rapid Mismatch Estimation (RME), an adaptive, controller-agnostic, probabilistic framework that estimates end-effector dynamics mismatches online, without relying on external force-torque sensors. From the robot's proprioceptive feedback, a Neural Network Model Mismatch Estimator generates a prior for a Variational Inference solver, which rapidly converges to the unknown parameters while quantifying uncertainty. With a real 7-DoF manipulator driven by a state-of-the-art passive impedance controller, RME adapts to sudden changes in mass and center of mass at the end-effector in $\sim400$ ms, in static and dynamic settings. We demonstrate RME in a collaborative scenario where a human attaches an unknown basket to the robot's end-effector and dynamically adds/removes heavy items, showcasing fast and safe adaptation to changing dynamics during physical interaction without any external sensory system.

In this paper, we tackle the problem of learning to play 3v3 multi-drone volleyball, a new embodied competitive task that requires both high-level strategic coordination and low-level agile control. The task is turn-based, multi-agent, and physically grounded, posing significant challenges due to its long-horizon dependencies, tight inter-agent coupling, and the underactuated dynamics of quadrotors. To address this, we propose Hierarchical Co-Self-Play (HCSP), a hierarchical reinforcement learning framework that separates centralized high-level strategic decision-making from decentralized low-level motion control. We design a three-stage population-based training pipeline to enable both strategy and skill to emerge from scratch without expert demonstrations: (I) training diverse low-level skills, (II) learning high-level strategy via self-play with fixed low-level controllers, and (III) joint fine-tuning through co-self-play. Experiments show that HCSP achieves superior performance, outperforming non-hierarchical self-play and rule-based hierarchical baselines with an average 82.9% win rate and a 71.5% win rate against the two-stage variant. Moreover, co-self-play leads to emergent team behaviors such as role switching and coordinated formations, demonstrating the effectiveness of our hierarchical design and training scheme.

We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.

Intelligent interaction with the real world requires robotic agents to jointly reason over high-level plans and low-level controls. This requirement is formalized in the task and motion planning (TAMP) problem, in which symbolic planning and continuous trajectory generation must be solved in a coordinated manner. Recently, foundation model-based approaches to TAMP have presented impressive results, including fast planning times and the execution of natural language instructions. Yet, the optimal interface between high-level plan and low-level motion generation remains to be found: prior approaches are limited by either too much abstraction (e.g., chaining simplified skill primitives) or a lack thereof (e.g., direct joint angle prediction). Our method introduces a novel technique employing a form of meta-optimization to address these shortcomings by: (i) using program search over trajectory optimization problems as an interface between foundation model and robot controllers, and (ii) leveraging a zero-order method to optimize numerical values in the foundation model output. Results on challenging object manipulation and drawing tasks confirm that our proposed method improves over prior TAMP approaches.

Humans efficiently generalize from limited demonstrations, but robots still struggle to transfer learned knowledge to complex, unseen tasks with longer horizons and increased complexity. We propose the first known method enabling robots to autonomously invent relational concepts directly from small sets of unannotated, unsegmented demonstrations. The learned symbolic concepts are grounded into logic-based world models, facilitating efficient zero-shot generalization to significantly more complex tasks. Empirical results demonstrate that our approach achieves performance comparable to hand-crafted models, successfully scaling execution horizons and handling up to 18 times more objects than seen in training, providing the first autonomous framework for learning transferable symbolic abstractions from raw robot trajectories.

End-to-end visuomotor policies trained using behavior cloning have shown a remarkable ability to generate complex, multi-modal low-level robot behaviors. However, at deployment time, these policies still struggle to act reliably when faced with out-of-distribution (OOD) visuals induced by objects, backgrounds, or environment changes. Prior works in interactive imitation learning solicit corrective expert demonstrations under the OOD conditions---but this can be costly and inefficient. We observe that task success under OOD conditions does not always warrant novel robot behaviors. In-distribution (ID) behaviors can directly be transferred to OOD conditions that share functional similarities with ID conditions. For example, behaviors trained to interact with in-distribution (ID) pens can apply to interacting with a visually-OOD pencil. The key challenge lies in disambiguating which ID observations functionally correspond to the OOD observation for the task at hand. We propose that an expert can provide this OOD-to-ID functional correspondence. Thus, instead of collecting new demonstrations and re-training at every OOD encounter, our method: (1) detects the need for feedback by checking if current observations are OOD and the most similar training observations show divergent behaviors (2) solicits functional correspondence feedback to disambiguate between those behaviors, and (3) intervenes on the OOD observations with the functionally corresponding ID observations to perform deployment-time generalization. We validate our method across diverse real-world robotic manipulation tasks with a Franka Panda robotic manipulator. Our results show that test-time functional correspondences can improve the generalization of a vision-based diffusion policy to OOD objects and environment conditions with low feedback.

We present Vision in Action (ViA), an active perception system for bimanual robot manipulation. ViA learns task-relevant active perceptual strategies (e.g., searching, tracking, and focusing) directly from human demonstrations. On the hardware side, ViA employs a simple yet effective 6-DoF robotic neck to enable flexible, human-like head movements. To capture human active perception strategies, we design a VR-based teleoperation interface that creates a shared observation space between the robot and the human operator. To mitigate VR motion sickness caused by latency in the robot's physical movements, the interface uses an intermediate 3D scene representation, enabling real-time view rendering on the operator side while asynchronously updating the scene with the robot's latest observations. Together, these design elements enable the learning of robust visuomotor policies for three complex, multi-stage bimanual manipulation tasks involving visual occlusions, significantly outperforming baseline systems.

Imitation learning has unlocked the potential for robots to exhibit highly dexterous behaviours. However, it still struggles with long-horizon, multi-object tasks due to poor sample efficiency and limited generalisation. Existing methods require a substantial number of demonstrations to cover possible task variations, making them costly and often impractical for real-world deployment. We address this challenge by introducing \emph{oriented affordance frames}, a structured representation for state and action spaces that improves spatial and intra-category generalisation and enables policies to be learned efficiently from only 10 demonstrations. More importantly, we show how this abstraction allows for compositional generalisation of independently trained sub-policies to solve long-horizon, multi-object tasks. To seamlessly transition between sub-policies, we introduce the notion of self-progress prediction, which we directly derive from the duration of the training demonstrations. We validate our method across three real-world tasks, each requiring multi-step, multi-object interactions. Despite the small dataset, our policies generalise robustly to unseen object appearances, geometries, and spatial arrangements, achieving high success rates without reliance on exhaustive training data. Video demonstration can be found on our anonymised project page: https://affordance-policy.github.io/.

Robot-assisted dressing has the potential to significantly improve the lives of individuals with mobility impairments. To ensure an effective and comfortable dressing experience, the robot must be able to handle challenging deformable garments, apply appropriate forces, and adapt to limb movements throughout the dressing process. Prior work often makes simplifying assumptions—such as static human limbs during dressing—which limits real-world applicability. In this work, we develop a robot-assisted dressing system capable of handling partial observations with visual occlusions, as well as robustly adapting to arm motions during the dressing process. Given a policy trained in simulation with partial observations, we propose a method to fine-tune it in the real world using a small amount of data and multi-modal feedback from vision and force sensing, to further improve the policy’s adaptability to arm motions and enhance safety. We evaluate our method in simulation with simplified articulated human meshes and in a real world human study with 12 participants across 264 dressing trials. Our policy successfully dresses two long-sleeve everyday garments onto the participants while being adaptive to various kinds of arm motions, and greatly outperforms prior baselines in terms of task completion and user feedback.

Visual-textual understanding is essential for language-guided robot manipulation. Recent works leverage pre-trained vision-language models to measure the similarity between encoded visual observations and textual instructions, and then train a model to map this similarity to robot actions. However, this two-step approach limits the model to capture the relationship between visual observations and textual instructions, leading to reduced precision in manipulation tasks. We propose to learn visual-textual associations through a self-supervised pretext task: reconstructing a masked goal image conditioned on an input image and textual instructions. This formulation allows the model to learn visual-action representations without robot action supervision. The learned representations can then be fine-tuned for manipulation tasks with only a few demonstrations. We also introduce the \textit{Omni-Object Pick-and-Place} dataset, which consists of annotated robot tabletop manipulation episodes, including 180 object classes and 3,200 instances with corresponding textual instructions. This dataset enables the model to acquire diverse object priors and allows for a more comprehensive evaluation of its generalisation capability across object instances. Experimental results on the five benchmarks, including both simulated and real-robot validations, demonstrate that our method outperforms prior art.