CoRL 2025 Papers

Learning visuamotor policy through imitation learning often suffers from perceptual challenges, where visual differences between training and evaluation environments degrade policy performance. Policies relying on state estimations like 6D pose, require task-specific tracking and are difficult to scale, while raw sensor-based policies may lack robustness to small visual disturbances. In this work, we leverage 2D keypoints — spatially consistent features in the image frame — as a state representation for robust policy learning, and apply it to both sim-to-real transfer and real-world imitation learning. However, the choice of which keypoints to use can vary across objects and tasks. We propose a novel method -ATK, to automatically select keypoints in a task-driven manner, such that the chosen keypoints are that are predictive of optimal behavior for the given task. Our proposal optimizes for a minimal set of task-relevant keypoints that preserve policy performance and robustness. We distill expert data (either from an expert policy in simulation or a human expert) into a policy that operates on RGB images while tracking the selected keypoints. By leveraging pre-trained visual modules, our system effectively tracks keypoints and transfers policies to the real-world evaluation scenario, even given perceptual challenges like transparent objects or fine-grained manipulation, or widely varying scene appearance. We validate our approach on various robotic tasks, demonstrating that these minimal keypoint representations improve robustness to visual disturbances and environmental variations.

Generalizable long-horizon robotic assembly requires reasoning at multiple levels of abstraction. While end-to-end imitation learning (IL) is a promising approach, it typically requires large amounts of expert demonstration data and often struggles to achieve the high precision demanded by assembly tasks. Reinforcement learning (RL) approaches, on the other hand, have shown some success in high-precision assembly, but suffer from sample inefficiency, which limits their effectiveness in long-horizon tasks. To address these challenges, we propose a hierarchical modular approach, named Adaptive Robotic Compositional Hierarchy (ARCH), which enables long-horizon, high-precision robotic assembly in contact-rich settings. ARCH employs a hierarchical planning framework, including a low-level primitive library of parameterized skills and a high-level policy. The low-level primitive library includes essential skills for assembly tasks, such as grasping and inserting. These primitives consist of both RL and model-based controllers. The high-level policy, learned via IL from a handful of demonstrations, without the need for teleoperation, selects the appropriate primitive skills and instantiates them with input parameters. We extensively evaluate our approach in simulation and on a real robotic manipulation platform. We show that ARCH generalizes well to unseen objects and outperforms baseline methods in terms of success rate and data efficiency.

Recent advances in generalist robot manipulation leverage pre-trained Vision–Language Models (VLMs) and large-scale robot demonstrations to tackle diverse tasks in a zero-shot manner. A key challenge remains: scaling high-quality, action-labeled robot demonstration data, which existing methods rely on for robustness and generalization. To address this, we propose a method that benefits from videos without action labels—featuring humans and/or robots in action—enhancing open-vocabulary performance and enabling data-efficient learning of new tasks. Our method extracts dense, dynamic 3D point clouds at the hand or gripper location and uses a proposed 3D dynamics predictor for self-supervision. This predictor is then tuned to an action predictor using a smaller labeled dataset for action alignment. We show that our method not only learns from unlabeled human and robot demonstrations—improving downstream generalist robot policies—but also enables robots to learn new tasks without action labels (i.e., out-of-action generalization) in both real-world and simulated settings.

Coordinated multi-robot navigation is essential for robots to operate as a team in diverse environments. During navigation, robot teams usually need to maintain specific formations, such as circular formations to protect human teammates at the center. However, in complex scenarios such as narrow corridors, rigidly preserving predefined formations can become infeasible. Therefore, robot teams must be capable of dynamically splitting into smaller subteams and adaptively controlling the subteams to navigate through such scenarios while preserving formations. To enable this capability, we introduce a novel method for SubTeaming and Adaptive Formation (STAF), which is built upon a unified hierarchical learning framework: (1) high-level deep graph cut for team splitting, (2) intermediate-level graph learning for facilitating coordinated navigation among subteams, and (3) low-level policy learning for controlling individual mobile robots to reach their goal positions while avoiding collisions. To evaluate STAF, we conducted extensive experiments in both indoor and outdoor environments using robotics simulations and physical robot teams. Experimental results show that STAF enables the novel capability for subteaming and adaptive formation control, and achieves promising performance in coordinated multi-robot navigation through challenging scenarios. More details are available on the project website: https://anonymous188.github.io/STAF/.

Exploration is crucial for legged robots to learn agile locomotion behaviors capable of overcoming diverse obstacles. For example, a robot may need to try different contact patterns and momentum profiles to successfully jump over an obstacle—but encouraging such diverse exploration is inherently challenging. As a result, training these behaviors often relies on additional techniques such as extensive reward engineering, expert demonstrations, or curriculum learning. However, these approaches limit generalizability, especially when prior knowledge or demonstration data is unavailable. In this work, we propose using unsupervised skill discovery as a skill-level exploration strategy to significantly reduce human engineering effort. Our learning framework enables the agent to autonomously discover diverse skills to overcome complex obstacles. To dynamically regulate the degree of exploration throughout training, we introduce a bi-level optimization process that learns a parameter to balance two distinct reward signals. We demonstrate that our method enables quadrupedal robots to acquire highly agile behaviors—including crawling, climbing, leaping, and complex maneuvers such as jumping off vertical walls. Finally, we successfully deploy the learned policy on real hardware, validating its transferability to the real world.

Tool use is essential for enabling robots to perform complex real-world tasks, but learning such skills requires extensive datasets. While teleoperation is widely used, it is slow, delay-sensitive, and poorly suited for dynamic tasks. In contrast, human videos provide a natural way for data collection without specialized hardware, though they pose challenges on robot learning due to viewpoint variations and embodiment gaps. To address these challenges, we propose a framework that transfers tool-use knowledge from humans to robots. To improve the policy's robustness to viewpoint variations, we use two RGB cameras to reconstruct 3D scenes and apply Gaussian splatting for novel view synthesis. We reduce the embodiment gap using segmented observations and tool-centric, task-space actions to achieve embodiment-invariant visuomotor policy learning. Our method achieves a 71\% improvement in task success and a 77\% reduction in data collection time compared to diffusion policies trained on teleoperation with equivalent time budgets. Our method also reduces data collection time by 41\% compared with the state-of-the-art data collection interface.

In Embodied Question Answering (EQA), agents must explore and develop a semantic understanding of an unseen environment in order to answer a situated question with confidence. This remains a challenging problem in robotics, due to the difficulties in obtaining useful semantic representations, updating these representations online, and leveraging prior world knowledge for efficient exploration and planning. Aiming to address these limitations, we propose GraphEQA, a novel approach that utilizes real-time 3D metric-semantic scene graphs (3DSGs) and task relevant images as multi-modal memory for grounding Vision-Language Models (VLMs) to perform EQA tasks in unseen environments. We employ a hierarchical planning approach that exploits the hierarchical nature of 3DSGs for structured planning and semantic-guided exploration. We evaluate GraphEQA in simulation on two benchmark datasets, HM-EQA and OpenEQA, and demonstrate that it outperforms key baselines by completing EQA tasks with higher success rates and fewer planning steps, and further demonstrate GraphEQA in two separate real world environments. Videos and code are available at https://grapheqa.github.io.

Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control---establishing a new paradigm for transparent and steerable foundation models in robotics.

We present DexUMI - a data collection and policy learning framework that uses the human hand as the natural interface to transfer dexterous manipulation skills to various robot hands. DexUMI incorporates hardware and software adaptations to minimize the embodiment gap between the human hand and various robot hands. The hardware adaptation bridges the kinematics gap with a wearable hand exoskeleton. It allows direct haptic feedback in manipulation data collection and adapts human motion to feasible robot hand motion. Our software adaptation bridges the visual gap by replacing the human hand in video data with high-fidelity robot hand inpainting. We demonstrate DexUMI's capabilities through comprehensive real-world experiments on two different dexterous robot hand hardware platforms, achieving an average task success rate of 86\%.

We introduce AgentWorld, an interactive simulation platform for developing household mobile manipulation capabilities. Our platform combines automated scene construction that encompasses layout generation, semantic asset placement, visual material configuration, and physics simulation, with a dual-mode teleoperation system supporting both wheeled bases and humanoid locomotion policies for data collection. The resulting AgentWorld Dataset captures diverse tasks ranging from primitive actions (pick-and-place, push-pull, etc.) to multistage activities (serve drinks, heat up food, etc.) across living rooms, bedrooms, and kitchens. Through extensive benchmarking of imitation learning methods including behavior cloning, action chunking transformers, diffusion policies, and vision-language-action models, we demonstrate the dataset's effectiveness for sim-to-real transfer. The integrated system provides a comprehensive solution for scalable robotic skill acquisition in complex home environments, bridging the gap between simulation-based training and real-world deployment.

We introduce ReWiND, a framework for learning robot manipulation tasks solely from language instructions without per-task demonstrations. Standard reinforcement learning (RL) and imitation learning methods require expert supervision through human-designed reward functions or demonstrations for every new task. In contrast, ReWiND starts from a small demonstration dataset to learn: (1) a data-efficient, language-conditioned reward function that labels the dataset with rewards, and (2) a language-conditioned policy pre-trained with offline RL using these rewards. Given an unseen task variation, ReWiND fine-tunes the pre-trained policy using the learned reward function, requiring minimal online interaction. We show that ReWiND’s reward model generalizes effectively to unseen tasks, outperforming baselines by up to 2.4X in reward generalization and policy alignment metrics. Finally, we demonstrate that ReWiND enables sample-efficient adaptation to new tasks in both simulation and on a real bimanual manipulation platform, taking a step towards scalable, real-world robot learning.

Quadrupedal robots have demonstrated remarkable agility and robustness in traversing complex terrains. However, they struggle with dynamic object interactions, where contact must be precisely sensed and controlled. To bridge this gap, we present LocoTouch, a system that equips quadrupedal robots with tactile sensing to address a particularly challenging task in this category: long-distance transport of unsecured cylindrical objects, which typically requires custom mounting or fastening mechanisms to maintain stability. For efficient large-area tactile sensing, we design a high-density distributed tactile sensor that covers the entire back of the robot. To effectively leverage tactile feedback for robot control, we develop a simulation environment with high-fidelity tactile signals, and train tactile-aware transport policies using a two-stage learning pipeline. Furthermore, we design a novel reward function to promote robust, symmetric, and frequency-adaptive locomotion gaits. After training in simulation, LocoTouch transfers zero-shot to the real world, reliably transporting a wide range of unsecured cylindrical objects with diverse sizes, weights, and surface properties. Moreover, it remains robust over long distances, on uneven terrain, and under severe perturbations.

Multi-object rearrangement is a challenging task that requires robots to reason about a physical 3D scene and the effects of a sequence of actions. While traditional task planning methods are shown to be effective for long-horizon manipulation, they require discretizing the continuous state and action space into symbolic descriptions of objects, object relationships, and actions. Our proposed method is instead able to take in a partially-observed point cloud observation of an initial scene and plan to a goal-satisfying configuration, without needing to discretize the set of actions or object relationships. To enable this, we formulate the planning problem as an A* search over the space of possible point cloud rearrangements. We sample point cloud transformations from a learned, domain-specific prior and then search for a sequence of such point cloud transformations that leads from the initial state to a goal. We evaluate our method in terms of task planning success and task execution success on a real-world, multi-step table bussing environment and a simulation block stacking environment. We experimentally demonstrate that our method produces successful plans and outperforms a policy-learning approach; we also perform ablations that show the importance of search in our approach.

Robot-assisted feeding requires reliable bite acquisition, a challenging task due to the complex interactions between utensils and food with diverse physical properties. These interactions are further complicated by the temporal variability of food properties—for example, steak becomes firm as it cools even during a meal. To address this, we propose SAVOR, a novel approach for learning skill affordances for bite acquisition—how suitable a manipulation skill (e.g., skewering, scooping) is for a given utensil-food interaction. In our formulation, skill affordances arise from the combination of tool affordances (what a utensil can do) and food affordances (what the food allows). Tool affordances are learned offline through calibration, where different utensils interact with a variety of foods to model their functional capabilities. Food affordances are characterized by physical properties such as softness, moisture, and viscosity, initially inferred through commonsense reasoning using a visually-conditioned language model and then dynamically refined through online multi-modal visuo-haptic perception using SAVOR-Net during interaction. Our method integrates these offline and online estimates to predict skill affordances in real time, enabling the robot to select the most appropriate skill for each food item. Evaluated on 20 single-item foods and 10 in-the-wild meals, our approach improves bite acquisition success by 13\% over state-of-the-art (SOTA) category-based methods (e.g. use skewer for fruits). These results highlight the importance of modeling interaction-driven skill affordances for generalizable and effective robot-assisted bite acquisition.

Humans and animals exhibit flexible locomotion strategies, such as learning to walk within minutes, and efficient adaptation to changes in morphology. In contrast, modern locomotion controllers are manually tuned for specific embodiments. In this paper, we present LocoFormer, a generalist policy that can control previously unseen legged and wheeled robots, even without precise knowledge of their kinematics. LocoFormer is able to adapt to changes in morphology and dynamics at test time. We find that two key choices enable adaptation. First, we train massive scale RL on procedurally generated robots with aggressive domain randomization. Second, in contrast to previous policies that are myopic with short context lengths, we extend context by orders of magnitude to span episode boundaries. We deploy the same LocoFormer to varied robots, and show robust control even with large disturbances such as weight and motor failures. In extreme scenarios, we see emergent adaptation across episodes, LocoFormer learns from falls in early episodes to improve control strategies in later ones. We believe this simple yet general recipe can be used to train foundation models for other robotic skills in the future. Videos at generalist-locomotion.github.io.

Terrain elevation modeling for off-road navigation aims to accurately estimate changes in terrain geometry in real-time and quantify the corresponding uncertainties. Having precise estimations and uncertainties plays a crucial role in planning and control algorithms to explore safe and reliable maneuver strategies. However, existing approaches, such as Gaussian Processes (GPs) and neural network-based methods, often fail to meet these needs. They are either unable to perform in real-time due to high computational demands, underestimating sharp geometry changes, or harming elevation accuracy when learned with uncertainties. Recently, Neural Processes (NPs) have emerged as a promising approach that integrates the Bayesian uncertainty estimation of GPs with the efficiency and flexibility of neural networks. Inspired by NPs, we propose an effective NP-based method that precisely estimates sharp elevation changes and quantifies the corresponding predictive uncertainty without losing elevation accuracy. Our method leverages semantic features from LiDAR and camera sensors to improve interpolation and extrapolation accuracy in unobserved regions. Also, we introduce a local ball-query attention mechanism to effectively reduce the computational complexity of global attention by 17\% while preserving crucial local and spatial information. We evaluate our method on off-road datasets having interesting geometric features, collected from trails, deserts, and hills. Our results demonstrate superior performance over baselines and showcase the potential of neural processes for effective and expressive terrain modeling in complex off-road environments.

Robotic manipulation in unstructured environments requires systems that can generalize across diverse tasks while maintaining robust and reliable performance. We introduce GVF-TAPE, a closed-loop framework that combines generative visual foresight with task-agnostic pose estimation to enable scalable robotic manipulation. GVF-TAPE employs a generative video model to predict future RGB-D frames from a single RGB side-view image and a task description, offering visual plans that guide robot actions. A decoupled pose estimation model then extracts end-effector poses from the predicted frames, translating them into executable commands via low-level controllers. By iteratively integrating video foresight and pose estimation in a closed loop, GVF-TAPE achieves real-time, adaptive manipulation across a broad range of tasks. Extensive experiments in both simulation and real-world settings demonstrate that our approach reduces reliance on task-specific action data and generalizes effectively, providing a practical and scalable solution for intelligent robotic systems

Scaling robot learning requires vast and diverse datasets. Yet the prevailing data collection paradigm—human teleoperation—remains costly and constrained by manual effort and physical robot access. We introduce Real2Render2Real (R2R2R), a novel approach for generating robot training data without relying on object dynamics simulation or teleoperation of robot hardware. The input is a smartphone-captured scan of one or more objects and a single video of a human demonstration. R2R2R renders thousands of high visual fidelity robot-agnostic demonstrations by reconstructing detailed 3D object geometry and appearance, and tracking 6-DoF object motion. R2R2R uses 3D Gaussian Splatting (3DGS) to enable flexible asset generation and trajectory synthesis for both rigid and articulated objects, converting these representations to meshes to maintain compatibility with scalable rendering engines like IsaacLab but with collision-modeling turned off. Robot demonstration data generated by R2R2R integrates directly with models that operate on robot proprioceptive states and image observations, such as vision-language-action models (VLA) and imitation learning policies. Physical experiments suggest that models trained on R2R2R data from a single human demonstration can match the performance of models trained on 150 human teleoperation demonstrations.

Sim-to-real discrepancies hinder learning-based policies from achieving high-precision tasks in the real world. While Domain Randomization (DR) is commonly used to bridge this gap, it often relies on heuristics and can lead to overly conservative policies with degrading performance when not properly tuned. System Identification (Sys-ID) offers a targeted approach, but standard techniques rely on differentiable dynamics and/or direct torque measurement, assumptions that rarely hold for contact-rich legged systems. To this end, we present SPI-Active (Sampling-based Parameter Identification with Active Exploration), a two-stage framework that estimates physical parameters of legged robots to minimize the sim-to-real gap. SPI-Active robustly identifies key physical parameters through massive parallel sampling, minimizing state prediction errors between simulated and real-world trajectories. To further improve the informativeness of collected data, we introduce an active exploration strategy that maximizes the Fisher Information of the collected real-world trajectories via optimizing the input commands of an exploration policy. This targeted exploration leads to accurate identification and better generalization across diverse tasks. Experimental results demonstrate that SPI-Active enables precise sim-to-real transfer of learned policies to the real world, outperforming baselines by 42-63% in various locomotion tasks. Videos at the anonymous website https://anonymous-spi-active.github.io/