CoRL 2025 Papers

We propose Hand-Eye Autonomous Delivery (HEAD), a framework that learns navigation, locomotion, and reaching skills for humanoids, directly from human motion and vision perception data. We take a modular approach where the high-level planner commands the target position and orientation of the hands and eyes of the humanoid, delivered by the low-level policy that controls the whole-body movements. Specifically, the low-level whole-body controller learns to track the three points (eyes, left hand, and right hand) from existing large-scale human motion capture data while high-level policy learns from human data collected by Aria glasses. Our modular approach decouples the ego-centric vision perception from physical actions, promoting efficient learning and scalability to novel scenes. We evaluate our method both in simulation and in the real-world, demonstrating humanoid's capabilities to navigate and reach in complex environments designed for humans.

Future robotic systems operating in real-world environments require on-board embodied intelligence without continuous cloud connection, balancing capabilities with constraints on computational power and memory. This work presents an extension of the R1-zero approach, which enables the usage of small parameter-count Large Language Models (LLMs) in the robotic domain. The R1-Zero approach was originally developed to enable mathematical reasoning in LLMs using static datasets. We extend it to the robotics domain through integration with a closed-loop Reinforcement Learning (RL) framework. This extension allows reasoning in Embodied Artificial Intelligence (EmbodiedAI) settings without relying solely on distillation of large models through Supervised Fine-Tuning (SFT). We show that small-scale LLMs can achieve effective reasoning performance by learning through closed-loop interaction with their environment, which enables tasks that previously required significantly larger models. A performance gain of 20.2% points over the SFT-based baseline is observed with a Qwen2.5-1.5B model. Using the proposed training procedure, Qwen2.5-3B achieves a 63.3% control adaptability score, surpassing the 58.5% obtained by the much larger, cloud-bound GPT-4o. These results highlight that practical, on-board deployment of small LLMs is not only feasible but can outperform larger models when trained through environmental interaction, underscoring the importance of an interactive, embodied learning framework for robotic EmbodiedAI — one grounded in practical experience rather than static supervision.

Traditional reinforcement learning lacks the ability to provide stability guarantees. More recent algorithms learn Lyapunov functions alongside the control policies to ensure stable learning. However, the current self-learned Lyapunov functions are sample inefficient due to their on-policy nature. This paper introduces a method for learning Lyapunov functions off-policy and incorporates the proposed off-policy Lyapunov function into the Soft Actor Critic and Proximal Policy Optimization algorithms to provide them with a data efficient stability certificate. Simulations of an inverted pendulum and a quadrotor illustrate the improved performance of the two algorithms when endowed with the proposed off-policy Lyapunov function.

Safe operation is essential for autonomous systems in safety-critical environments such as urban air mobility. Value function-based safety filters provide formal guarantees on safety, wrapping learned or planning-based controllers with a layer of protection. Recent approaches leverage offline learned value functions to scale these safety filters to high-dimensional systems. Yet these methods assume detailed prior knowledge of all possible sources of model mismatch, in the form of disturbances, in the environment -- information that is typically unavailable in real world settings. Even in well-mapped environments like urban canyons or industrial sites, drones encounter complex, spatially-varying disturbances arising from payload-drone interaction, turbulent airflow, and other environmental factors. We introduce Space2Time, which enables safe and adaptive deployment of offline-learned safety filters under unknown, spatially-varying disturbances. The key idea is to reparameterize spatial disturbances as a time-varying formulation, allowing the use of temporally varying precomputed value functions during online operation. We validate Space2Time through extensive simulations on diverse quadcopter models and real-world hardware experiments, demonstrating significantly improved safety performance over worst-case and naive baselines.

Collecting demonstrations enriched with fine-grained tactile information is critical for dexterous manipulation, particularly in contact-rich tasks that require precise force control and physical interaction. While prior works primarily focus on teleoperation or video-based retargeting, they often suffer from kinematic mismatches and the absence of real-time tactile feedback, hindering the acquisition of high-fidelity tactile data. To mitigate this issue, we propose KineDex, a hand-over-hand kinesthetic teaching paradigm in which the operator’s motion is directly transferred to the dexterous hand, enabling the collection of physically grounded demonstrations enriched with accurate tactile feedback. To resolve occlusions from human hand, we apply inpainting technique to preprocess the visual observations. Based on these demonstrations, we then train a visuomotor policy using tactile-augmented inputs and implement force control during deployment for precise contact-rich manipulation. We evaluate KineDex on a suite of challenging contact-rich manipulation tasks, including particularly difficult scenarios such as squeezing toothpaste onto a toothbrush, which require precise multi-finger coordination and stable force regulation. Across these tasks, KineDex achieves an average success rate of 74.4%, representing a 57.7% improvement over the variant without force control. Comparative experiments with teleoperation and user studies further validate the advantages of KineDex in data collection efficiency and operability. Specifically, KineDex collects data over twice as fast as teleoperation across two tasks of varying difficulty, while maintaining a near-100% success rate, compared to under 50% for teleoperation.

Embodied visual tracking is a fundamental skill in Embodied AI, enabling an agent to follow a specific target in dynamic environments using only egocentric vision. This task is inherently challenging as it requires both accurate target recognition and effective trajectory planning under conditions of severe occlusion and high scene dynamics. Existing approaches typically address this challenge through a modular separation of recognition and planning. In this work, we propose TrackVLA, a Vision-Language-Action (VLA) model that learns the synergy between object recognition and trajectory planning. Leveraging a shared LLM backbone, we employ a language modeling head for recognition and an anchor-based diffusion model for trajectory planning. To train TrackVLA, we construct an Embodied Visual Tracking Benchmark (EVT-Bench) and collect diverse difficulty levels of recognition samples, resulting in a dataset of 1.7 million samples. Through extensive experiments in both synthetic and real-world environments, TrackVLA demonstrates SOTA performance and strong generalizability. It significantly outperforms existing methods on public benchmarks in a zero-shot manner while remaining robust to high dynamics and occlusion in real-world scenarios at 10 FPS inference speed.

Learning-based robotics research driven by data demands a new approach to robot hardware design—one that serves as both a platform for policy execution and a tool for embodied data collection. We introduce ToddlerBot, a low-cost, open-source humanoid robot platform designed for robotics and AI research. ToddlerBot enables seamless acquisition of high-quality simulation and real-world data. The plug-and-play zero-point calibration and transferable motor system identification ensure a high-fidelity digital twin and zero-shot sim-to-real policy transfer. A user-friendly teleoperation interface streamlines real-world data collection from human demonstrations. With its data collection ability and anthropomorphic design, ToddlerBot is ideal for whole-body loco-manipulation research. Additionally, ToddlerBot's compact size (0.56 m, 3.4 kg) ensures safe operation in real-world environments. Reproducibility is achieved with entirely 3D-printed, open-source design and off-the-shelf components, keeping the total cost under 6,000 USD. This allows assembly and maintenance with basic technical expertise, as validated by successful independent replications of the system. We demonstrate ToddlerBot's capabilities through arm span, payload, endurance tests, loco-manipulation tasks, and a collaborative long-horizon scenario where two robots tidy a toy session together. By advancing ML-compatibility, capability, and reproducibility, ToddlerBot provides a robust and scalable platform for policy learning and execution in robotics research.

Diffusion Policy (DP) enables robots to learn complex behaviors by imitating expert demonstrations through action diffusion. However, in practical applications, hardware limitations often degrade data quality, while real-time constraints restrict model inference to instantaneous state and scene observations. These limitations seriously reduce the efficacy of learning from expert demonstrations, resulting in failures in object localization, grasp planning, and long-horizon task execution. To address these challenges, we propose Causal Diffusion Policy (CDP), a novel transformer-based diffusion model that enhances action prediction by conditioning on historical action sequences, thereby enabling more coherent and context-aware visuomotor policy learning. In practice, to further mitigate the computational cost associated with autoregressive inference, a caching mechanism is also introduced to store attention key-value pairs from previous timesteps, substantially reducing redundant computations during execution. Extensive experiments in both simulated and real-world environments, spanning diverse 2D and 3D manipulation tasks, demonstrate that CDP uniquely leverages historical action sequences to achieve significantly higher accuracy than existing methods. Moreover, even when faced with degraded observation quality, CDP maintains remarkable precision by reasoning through temporal continuity, which highlights its practical robustness for robotic control under realistic, imperfect conditions.

Mobile robots navigating in crowds trained using reinforcement learning are known to suffer performance degradation when faced with out-of-distribution scenarios. We propose that by properly accounting for the uncertainties of pedestrians, a robot can learn safe navigation policies that are robust to distribution shifts. Our method augments agent observations with prediction uncertainty estimates generated by adaptive conformal inference, and it uses these estimates to guide the agent’s behavior through constrained reinforcement learning. The system helps regulate the agent’s actions and enables it to adapt to distribution shifts. In the in-distribution setting, our approach achieves a 96.93% success rate, which is over 8.80% higher than the previous state-of-the-art baselines with over 3.72 times fewer collisions and 2.43 times fewer intrusions into ground-truth human future trajectories. In three out-of-distribution scenarios, our method shows much stronger robustness when facing distribution shifts in velocity variations, policy changes, and transitions from individual to group dynamics. We deploy our method on a real robot, and experiments show that the robot makes safe and robust decisions when interacting with both sparse and dense crowds.

The human body demonstrates exceptional motor capabilities—such as standing steadily on one foot or performing a high kick with the leg raised over 1.5 meters—both requiring precise balance control. While recent research on humanoid control has leveraged reinforcement learning to track human motions for skill acquisition, applying this paradigm to balance-intensive tasks remains challenging. In this work, we identify three key obstacles: instability from reference motion errors, learning difficulties due to morphological mismatch, and the sim-to-real gap caused by sensor noise and unmodeled dynamics. To address these challenges, we propose $\textbf{HuB}$ ($\textbf{Hu}$manoid $\textbf{B}$alance), a unified framework that integrates $\textit{reference motion refinement}$, $\textit{balance-aware policy learning}$, and $\textit{sim-to-real robustness training}$, with each component targeting a specific challenge. We validate our approach on the Unitree G1 humanoid robot across challenging quasi-static balance tasks, including extreme single-legged poses such as $\texttt{Swallow Balance}$ and $\texttt{Bruce Lee’s Kick}$. Our policy remains stable even under strong physical disturbances—such as a forceful soccer strike—while baseline methods consistently fail to complete these tasks.

We introduce RoLA, a framework that transforms any in‑the‑wild image into an interactive, physics‑enabled robotic environment. Unlike previous methods, RoLA operates directly on a single image without requiring additional hardware or digital assets. Our framework democratizes robotic data generation by producing massive visuomotor robotic demonstrations within minutes from a wide range of image sources, including camera captures, robotic datasets, and Internet images. At its core, our approach combines a novel method for single-view physical scene recovery with an efficient visual blending strategy for photorealistic data collection. We demonstrate RoLA's versatility across applications like scalable robotic data generation and augmentation, robot learning from internet images, and single-image real-to-sim-to-real systems for manipulators and humanoids. Video results are available at our \href{https://rola-2025.github.io/}{project page}.

This paper considers the problem of enabling robots to navigate dynamic environments while following instructions. The challenge lies in the combinatorial nature of instruction specifications: each instruction can include multiple specifications, and the number of possible specification combinations grows exponentially as the robot’s skill set expands. For example, “overtake the pedestrian while staying on the right side of the road” consists of two specifications: *"overtake the pedestrian"* and *"walk on the right side of the road."* To tackle this challenge, we propose ComposableNav, based on the intuition that following an instruction involves independently satisfying its constituent specifications, each corresponding to a distinct motion primitive. Using diffusion models, ComposableNav learns each primitive separately, then composes them in parallel at deployment time to satisfy novel combinations of specifications unseen in training. Additionally, to avoid the onerous need for demonstrations of individual motion primitives, we propose a two-stage training procedure: (1) supervised pre-training to learn a base diffusion model for dynamic navigation, and (2) reinforcement learning fine-tuning that molds the base model into different motion primitives. Through simulation and real-world experiments, we show that ComposableNav enables robots to follow instructions by generating trajectories that satisfy diverse and unseen combinations of specifications, significantly outperforming both non-compositional VLM-based policies and costmap composing baselines.

Mimicry is a fundamental learning mechanism in humans, enabling individuals to learn new tasks by observing and imitating experts. However, applying this ability to robots presents significant challenges due to the inherent differences between human and robot embodiments in both their visual appearance and physical capabilities. While previous methods bridge this gap using cross-embodiment datasets with shared scenes and tasks, collecting such aligned data between humans and robots at scale is not trivial. In this paper, we propose UniSkill, a novel framework that learns embodiment-agnostic skill representations from large-scale cross-embodiment video data without any labels, enabling skills extracted from human video prompts to effectively transfer to robot policies trained only on robot data. Our experiments in both simulation and real-world environments show that our cross-embodiment skills successfully guide robots in selecting appropriate actions, even with unseen video prompts.

Reinforcement learning (RL) has made significant strides in legged robot control, enabling locomotion across diverse terrains and complex loco-manipulation capabilities. However, the commonly used position or velocity tracking-based objectives are agnostic to forces experienced by the robot, leading to stiff and potentially dangerous behaviors and poor control during forceful interactions. To address this limitation, we present Force-Adaptive Control via Impedance Reference Tracking (FACET). Inspired by impedance control, we use RL to train a control policy to imitate a virtual mass-spring-damper system, allowing fine-grained control under external forces by manipulating the virtual spring. In simulation, we demonstrate that our quadruped robot achieves improved robustness to large impulses (up to 200 Ns) and exhibits controllable compliance, achieving an 80\% reduction in collision impulse. The policy is deployed to a physical robot, demonstrating both compliant behavior, such as initiation/cessation of movement with finger tip, and the ability to pull payloads up to 10kg. Further extension to a legged loco-manipulator and a humanoid shows the applicability of our method to more complex settings to enable whole-body compliance control.

RGB-based novel object pose estimation is critical for rapid deployment in robotic applications, yet zero-shot generalization remains a key challenge. In this paper, we introduce PicoPose, a novel framework designed to tackle this task using a three-stage pixel-to-pixel correspondence learning process. Firstly, PicoPose matches features from the RGB observation with those from rendered object templates, identifying the best-matched template and establishing coarse correspondences. Secondly, PicoPose smooths the correspondences by globally regressing a 2D affine transformation, including in-plane rotation, scale, and 2D translation, from the coarse correspondence map. Thirdly, PicoPose applies the affine transformation to the feature map of the best-matched template and learns correspondence offsets within local regions to achieve fine-grained correspondences. By progressively refining the correspondences, PicoPose significantly improves the accuracy of object poses computed via PnP/RANSAC. PicoPose achieves state-of-the-art performance on the seven core datasets of the BOP benchmark, demonstrating exceptional generalization to novel objects. Our code and models will be made publicly available.

Offline Imitation Learning (IL) methods such as Behavior Cloning are effective at acquiring complex robotic manipulation skills. However, existing IL-trained policies are confined to execute the task at the same speed as shown in demonstration data. This limits the task throughput of a robotic system, a critical requirement for applications such as industrial automation. In this paper, we introduce and formalize the novel problem of enabling faster-than-demonstration execution of visuomotor policies and identify fundamental challenges in robot dynamics and state-action distribution shifts. We instantiate the key insights as SAIL (Speed Adaptation for Imitation Learning), a full-stack system integrating four tightly-connected components: (1) a consistency-preserving action inference algorithm for smooth motion at high speed, (2) high-fidelity tracking of controller-invariant motion targets, (3) adaptive speed modulation that dynamically adjusts execution speed based on motion complexity, and (4) action scheduling to handle real-world system latencies. Experiments on 12 tasks across simulation and two real, distinct robot platforms shows that SAIL achieves up to a {4$\times$ speedup} over demonstration speed in simulation and up to {3.2$\times$ speedup} in the real world. Additional detail is available at https://sail-robot.github.io

For robot manipulation, both the controller and end-effector design are crucial. Compared with rigid grippers, soft grippers are more generalizable by deforming to different geometries, but designing such a gripper and finding its grasp pose remains challenging. In this paper, we propose a co-design framework that generates an optimized soft gripper’s block-wise stiffness distribution and its grasping pose, using a neural physics model trained in simulation. We adopt a uniform-pressure tendon model, then generate a diverse dataset by randomizing both gripper pose and design parameters. A neural network is trained to approximate this forward simulation, yielding a fast, differentiable surrogate. We embed that surrogate in an end-to-end optimization loop to recover the ideal stiffness configuration and best grasp pose. Finally, we 3D-print the optimized grippers of various stiffness by changing the printing infills and parameters. We demonstrate that our co-designed grippers significantly outperform baseline designs in terms of force closure and success rate.

Model-free diffusion planners have shown great promise for robot motion planning, but practical robotic systems often require combining them with model-based optimization modules to enforce constraints, such as safety. Na\"ively integrating these modules presents compatibility challenges when diffusion's multi-modal outputs behave adversarially to optimization-based modules. To address this, we introduce Joint Model-based Model-free Diffusion (JM2D), a novel generative modeling framework. JM2D formulates module integration as a joint sampling problem to maximize compatibility via an interaction potential, without additional training. Using importance sampling, JM2D guides modules outputs based only on evaluations of the interaction potential, thus handling non-differentiable objectives commonly arising from non-convex optimization modules. We evaluate JM2D via application to aligning diffusion planners with safety modules on offline RL and robot manipulation. JM2D significantly improves task performance compared to conventional safety filters without sacrificing safety. Further, we show that conditional generation is a special case of JM2D and elucidate key design choices by comparing with SOTA gradient-based and projection-based diffusion planners. More details at: \url{https://sites.google.com/view/joint-mbmf-diffusion}

End-to-end imitation learning offers a promising approach for training robot policies. However, generalizing to new settings—such as unseen scenes, tasks, and object instances—remains a significant challenge. Although large-scale robot demonstration datasets have shown potential for inducing generalization, they are resource-intensive to scale. In contrast, human video data is abundant and diverse, presenting an attractive alternative. Yet, these human-video datasets lack action labels, complicating their use in imitation learning. Existing methods attempt to extract grounded action representations (e.g., hand poses), but resulting policies struggle to bridge the embodiment gap between human and robot actions. We propose an alternative approach: leveraging language-based reasoning from human videos—essential for guiding robot actions—to train generalizable robot policies. Building on recent advances in reasoning-based policy architectures, we introduce Reasoning through Action-free Data (RAD). RAD learns from both robot demonstration data (with reasoning and action labels) and action-free human video data (with only reasoning labels). The robot data teaches the model to map reasoning to low-level actions, while the action-free data enhances reasoning capabilities. Additionally, we release a new dataset of 3,377 human-hand demonstrations compatible with the Bridge V2 benchmark. This dataset includes chain-of-thought reasoning annotations and hand-tracking data, and is aimed at facilitating future research on reasoning-driven robot learning. Our experiments demonstrate that RAD enables effective transfer across the embodiment gap, allowing robots to perform tasks seen only in action-free data. Furthermore, scaling up action-free reasoning data significantly improves policy performance and generalization to novel tasks. These results highlight the promise of reasoning-driven learning from action-free datasets for advancing generalizable robot control. See website with videos: https://rad-generalization-s.github.io/