CoRL 2025 论文列表

Unsupervised Skill Discovery (USD) allows agents to autonomously learn diverse behaviors without task-specific rewards. While recent USD methods have shown promise, their application to real-world robotics remains underexplored. In this paper, we propose a modular USD framework to address the challenges in safety, interpretability, and deployability of the learned skills. Our approach factorizes the state space to learn disentangled skill representations and assigns different skill discovery algorithms to each factor based on the desired intrinsic reward function. To encourage structured morphology-aware skills, we introduce symmetry-based inductive biases tailored to individual factors. We also incorporate a style factor and regularization penalties to promote safe and robust behaviors. We evaluate our framework in simulation using a quadrupedal robot and demonstrate zero-shot transfer of the learned skills to real hardware. Our results show that factorization and symmetry lead to the discovery of structured, human-interpretable behaviors, while the style factor and penalties enhance safety and diversity. Additionally, we show that the learned skills can be used for downstream tasks and perform on par with oracle policies trained with hand-crafted rewards. To facilitate future research, we will release our code upon publication.

LiDAR place recognition is a critical capability for autonomous navigation and cross-modal localization in large-scale outdoor environments. Existing approaches predominantly depend on pre-built 3D dense maps or aerial imagery, which impose significant storage overhead and lack real-time adaptability. In this paper, we propose OPAL, a novel network for LiDAR place recognition that leverages OpenStreetMap (OSM) as a lightweight and up-to-date prior. Our key innovation lies in bridging the domain disparity between sparse LiDAR scans and structured OSM data through two carefully designed components. First, a cross-modal visibility mask that identifies maximal observable regions from both modalities to guide feature learning. Second, an adaptive radial fusion module that dynamically consolidates radial features into discriminative global descriptors. Extensive experiments on the KITTI and KITTI-360 datasets demonstrate OPAL’s superiority, achieving 15.98% higher recall at @1m threshold for top-1 retrieved matches, along with 12x faster inference speed compared to the state-of-the-art approach. Code and datasets will be publicly available.

Reliable scene understanding requires not only accurate predictions but also well-calibrated confidence estimates to ensure calibrated uncertainty estimation, especially in safety-critical domains like autonomous driving. In this context, semantic segmentation of LiDAR points supports real-time 3D scene understanding, where reliable uncertainty estimates help identify potentially erroneous predictions. While most existing calibration approaches focus on modeling epistemic uncertainty, they often overlook aleatoric uncertainty arising from measurement inaccuracies, which is especially prevalent in LiDAR data and essential for real-world deployment. In this work, we introduce a sampling-free approach for estimating well-calibrated confidence values by explicitly modeling aleatoric uncertainty in semantic segmentation, achieving alignment with true classification accuracy and reducing inference time compared to sampling-based methods. Evaluated on the real-world SemanticKITTI benchmark, our approach achieves 1.70\% and 1.33\% Adaptive Calibration Error (ACE) in semantic segmentation of LiDAR data using RangeViT and SalsaNext models, and is more than one order of magnitude faster than the comparable baseline. Furthermore, reliability diagrams reveal that our method produces underconfident rather than overconfident predictions — an advantageous property in safety-critical systems.

Teaching robots dexterous manipulation skills often requires collecting hundreds of demonstrations using wearables or teleoperation, a process that is challenging to scale. Videos of human-object interactions are easier to collect and scale, but leveraging them directly for robot learning is difficult due to the lack of explicit action labels and human-robot embodiment differences. We propose Human2Sim2Robot, a novel real-to-sim-to-real framework for training dexterous manipulation policies using only one RGB-D video of a human demonstrating a task. Our method utilizes reinforcement learning (RL) in simulation to cross the embodiment gap without relying on wearables, teleoperation, or large-scale data collection. From the video, we extract: (1) the object pose trajectory to define an object-centric, embodiment-agnostic reward, and (2) the pre-manipulation hand pose to initialize and guide exploration during RL training. These components enable effective policy learning without any task-specific reward tuning. In the single human demo regime, Human2Sim2Robot outperforms object-aware replay by over 55% and imitation learning by over 68% on grasping, non-prehensile manipulation, and multi-step tasks. Website: https://human2sim2robot.github.io

Recent advances in generative world models have enabled classical safe control methods, such as Hamilton-Jacobi (HJ) reachability, to generalize to complex robotic systems operating directly from high-dimensional sensor observations. However, obtaining comprehensive coverage of all safety-critical scenarios during world model training is extremely challenging. As a result, latent safety filters built on top of these models may miss novel hazards and even fail to prevent known ones, overconfidently misclassifying risky out-of-distribution (OOD) situations as safe. To address this, we introduce an uncertainty-aware latent safety filter that proactively steers robots away from both known and unseen failures. Our key idea is to use the world model’s epistemic uncertainty as a proxy for identifying unseen potential hazards. We propose a principled method to detect OOD world model predictions by calibrating an uncertainty threshold via conformal prediction. By performing reachability analysis in an augmented state space—spanning both the latent representation and the epistemic uncertainty—we synthesize a latent safety filter that can reliably safeguard arbitrary policies from both known and unseen safety hazards. In simulation and hardware experiments on vision-based control tasks with a Franka manipulator, we show that our uncertainty-aware safety filter preemptively detects potential unsafe scenarios and reliably proposes safe, in-distribution actions.

Human skin provides a rich tactile sensing stream, localizing intentional and unintentional contact events over a large and contoured region. Replicating these tactile sensing capabilities for dexterous robotic manipulation systems remains a longstanding challenge. In this work, we take a step towards this goal by introducing DexSkin. DexSkin is a soft, conformable capacitive electronic skin that enables sensitive, localized, and calibratable tactile sensing, and can be tailored to varying geometries. We demonstrate its efficacy for learning downstream robotic manipulation by sensorizing a pair of parallel jaw gripper fingers, providing tactile coverage across almost the entire finger surfaces. We empirically evaluate DexSkin’s capabilities in learning challenging manipulation tasks that require sensing coverage across the entire surface of the fingers, such as reorienting objects in hand and wrapping elastic bands around boxes, in a learning-from-demonstration framework. We then show that, critically for data-driven approaches, DexSkin can be calibrated to enable model transfer across sensor instances, and demonstrate its applicability to online reinforcement learning on real robots. Our results highlight DexSkin’s suitability and practicality for learning real-world, contact-rich manipulation.

Imitating tool manipulation from human videos offers an intuitive approach to teaching robots, while also providing a promising and scalable alternative to labor-intensive teleoperation data collection for visuomotor policy learning. While humans can mimic tool manipulation behavior by observing others perform a task just once and effortlessly transfer the skill to diverse tools for functionally equivalent tasks, current robots struggle to achieve this level of generalization. A key challenge lies in establishing function-level correspondences, considering the significant geometric variations among functionally similar tools, referred to as intra-function variations. To address this challenge, we propose MimicFunc, a framework that establishes functional correspondences with function frame, a function-centric local coordinate frame constructed with 3D functional keypoints, for imitating tool manipulation skills. Experiments demonstrate that MimicFunc effectively enables the robot to generalize the skill from a single RGB-D human video to manipulating novel tools for functionally equivalent tasks. Furthermore, leveraging MimicFunc's one-shot generalization capability, the generated rollouts can be used to train visuomotor policies without requiring labor-intensive teleoperation data collection for novel objects.

How can we teach humanoids to climb staircases and sit on chairs using the surrounding environment context? Arguably the simplest way is to _just show them_—casually capture a human motion video and feed it to humanoids. We introduce **VideoMimic**, a real-to-sim-to-real pipeline that mines everyday videos, jointly reconstructs the humans and the environment, and produces whole-body control policies for humanoid robots that perform the corresponding skills. We demonstrate the results of our pipeline on real humanoid robots, showing robust, repeatable contextual control such as staircase ascents and descents, sitting and standing from chairs and benches, as well as other dynamic whole-body skills all from a single policy, conditioned on the environment and global root commands. We hope our data and approach help enable a scalable path towards teaching humanoids to operate in diverse real-world environments.

Humanoid robot teleoperation plays a vital role in demonstrating and collecting data for complex interactions. Current methods suffer from two key limitations: (1) restricted controllability due to decoupled upper- and lower-body control, and (2) severe drift caused by open-loop execution. These issues prevent humanoid robots from performing coordinated whole-body motions required for long-horizon loco-manipulation tasks. We introduce CLONE, a whole-body teleoperation system that overcomes these challenges through three key contributions: (1) a Mixture-of-Experts (MoE) whole-body control policy that enables complex coordinated movements, such as “picking up an object from the ground” and “placing it in a distant bin”; (2) a closed-loop error correction mechanism using LiDAR odometry, reducing translational drift to 12cm over 8.9-meter trajectories; and (3) a systematic data augmentation strategy that ensures robust performance under diverse, previously unseen operator poses. In extensive experiments, CLONE demonstrates robust performance across diverse scenarios while maintaining stable whole-body control. These capabilities significantly advance humanoid robotics by enabling the collection of long-horizon interaction data and establishing a foundation for more sophisticated humanoid-environment interaction in both research and practical applications.

Can your humanoid walk up and hand you a full cup of beer—without spilling a drop? While humanoids are increasingly featured in flashy demos—dancing, delivering packages, traversing rough terrain—fine-grained control during locomotion remains a significant challenge. In particular, stabilizing a filled end-effector (EE) while walking is far from solved, due to a fundamental mismatch in task dynamics: locomotion demands slow-timescale, robust control, whereas EE stabilization requires rapid, high-precision corrections. To address this, we propose SoFTA, a Slow-Fast Two-Agent framework that decouples upper-body and lower-body control into separate agents operating at different frequencies and with distinct rewards. This temporal and objective separation mitigates policy interference mitagates objective conflict and enables coordinated whole-body behavior. SoFTA executes upper-body actions at 100 Hz for precise EE control and lower-body actions at 50 Hz for robust gait. It reduces EE acceleration by 2-5x to baselines and performs 2–3x closer to human-level stability, enabling delicate tasks such as carrying nearly full cups, capturing steady video during locomotion, and disturbance rejection with EE stability.

Effective policy learning for robotic manipulation requires scene representations that selectively capture task-relevant environmental features. Current approaches typically employ task-agnostic representation extraction, failing to emulate the dynamic perceptual adaptation observed in human cognition. We present HyperTASR, a hypernetwork-driven framework that modulates scene representations based on both task objectives and the execution phase. Our architecture dynamically generates representation transformation parameters conditioned on task specifications and progression state, enabling representations to evolve contextually throughout task execution. This approach maintains architectural compatibility with existing policy learning frameworks while fundamentally reconfiguring how visual features are processed. Unlike methods that simply concatenate or fuse task embeddings with task-agnostic representations, HyperTASR establishes computational separation between task-contextual and state-dependent processing paths, enhancing learning efficiency and representational quality. Comprehensive evaluations in both simulation and real-world environments demonstrate substantial performance improvements across different representation paradigms. Most notably, HyperTASR elevates success rates by over 27\% when applied to GNFactor and achieves unprecedented single-view performance exceeding 80\% success with 3D Diffuser Actor. Through ablation studies and attention visualization, we confirm that our approach selectively prioritizes task-relevant scene information, closely mirroring human adaptive perception during manipulation tasks.

Training general-purpose robots requires learning from large and diverse data sources. Current approaches rely heavily on teleoperated demonstrations which are difficult to scale. We present a scalable framework for training manipulation policies directly from human video demonstrations, requiring no robot data. Our method converts human demonstrations into robot-compatible observation-action pairs using hand pose estimation and visual data editing. We inpaint the human arm and overlay a rendered robot to align the visual domains. This enables zero-shot deployment on real hardware without any fine-tuning. We demonstrate strong success rates—up to 92%—on a range of tasks including deformable object manipulation, multi-object sweeping, and insertion. Our approach generalizes to novel environments and supports closed-loop execution. By demonstrating that effective policies can be trained using only human videos, our method broadens the path to scalable robot learning. Videos are available at https://phantom-training-robots.github.io.

Human videos offer a scalable way to train robot manipulation policies, but lack the action labels needed by standard imitation learning algorithms. Existing cross-embodiment approaches try to map human motion to robot actions, but often fail when the embodiments differ significantly. We propose X-Sim, a real-to-sim-to-real framework that uses object motion as a dense and transferable signal for learning robot policies. X-Sim starts by reconstructing a photorealistic simulation from an RGBD human video and tracking object trajectories to define object-centric rewards. These rewards are used to train a reinforcement learning (RL) policy in simulation. The learned policy is then distilled into an image-conditioned diffusion policy using synthetic rollouts rendered with varied viewpoints and lighting. To transfer to the real world, X-Sim introduces an online domain adaptation technique that aligns real and simulated observations during deployment. Importantly, X-Sim does not require any robot teleoperation data. We evaluate it across 5 manipulation tasks in 2 environments and show that it: (1) improves task progress by 30\% on average over hand-tracking and sim-to-real baselines, (2) matches behavior cloning with 10x less data collection, and (3) generalizes to new camera viewpoints and test-time changes.

Legged robots have the potential to leverage obstacles to climb steep sand slopes. However, efficiently repositioning these obstacles to desired locations is challenging. Here we present DiffusiveGRAIN, a learning-based method that enables a multi-legged robot to strategically induce localized sand avalanches during locomotion and indirectly manipulate obstacles. We conducted 375 trials, systematically varying obstacle spacing, robot orientation, and leg actions in 75 of them. Results show that movement of closely-spaced obstacles exhibit significant interference, requiring joint modeling. In addition, different multi-leg excavation actions could cause distinct robot state changes, necessitating integrated planning of manipulation and locomotion. To address these challenges, DiffusiveGRAIN includes a diffusion-based environment predictor to capture multi-obstacle movements under granular flow interferences and a robot state predictor to estimate changes in robot state from multi-leg action patterns. Deployment experiments (90 trials) demonstrate that by integrating the environment and robot state predictors, the robot can autonomously plan its movements based on loco-manipulation goals, successfully shifting closely located rocks to desired locations in over 65% of trials. Our study showcases the potential for a locomoting robot to strategically manipulate obstacles to achieve improved mobility on challenging terrains.

3D articulated objects modeling has long been a challenging problem, since it requires to capture both accurate surface geometries and semantically meaningful and spatially precise structures, parts, and joints. Existing methods heavily depend on training data from a limited set of handcrafted articulated object categories (e.g., cabinets and drawers), which restricts their ability to model a wide range of articulated objects in an open-vocabulary context. To address these limitations, we propose Articulate AnyMesh, an automated framework that is able to convert any rigid 3D mesh into its articulated counterpart in an open-vocabulary manner. Given a 3D mesh, our framework utilizes advanced Vision-Language Models and visual prompting techniques to extract semantic information, allowing for both the segmentation of object parts and the construction of functional joints. Our experiments show that Articulate AnyMesh can generate large-scale, high-quality 3D articulated objects, including tools, toys, mechanical devices, and vehicles, significantly expanding the coverage of existing 3D articulated object datasets. Additionally, we show that these generated assets can facilitate the acquisition of new articulated object manipulation skills in simulation, which can then be transferred to a real robotic system.

In this work, we study the problem of data retrieval for few-shot imitation learning: select data from a large dataset to train a performant policy for a specific task, given only a few target demonstrations. Prior methods retrieve data using a single-feature distance heuristic, assuming that the best demonstrations are those that most closely resemble the target examples in visual, semantic, or motion space. However, this approach captures only a subset of the relevant information and is prone to introducing detrimental demonstrations, e.g., retrieving data from unrelated tasks due to similar scene layouts, or selecting similar motions from tasks with divergent goals. We present COLLAGE, a method for COLLective data AGgrEgation in few-shot imitation learning that uses an adaptive late fusion mechanism to guide the selection of relevant demonstrations based on a task12 specific combination of multiple cues. COLLAGE follows a simple, but flexible and efficient data aggregation recipe: it assigns weights to subsets of the dataset that are pre-selected using a single feature (e.g., appearance, shape, or language similarity), based on their task relevance, measured by how well a policy trained on each subset predicts actions in the few target demonstrations. These weights are then used during policy training to perform importance sampling over the aggregated dataset, sampling data more densely or sparsely, according to their estimated relevance. This weighted aggregation strategy is general and feature-agnostic, allowing COLLAGE to combine and leverage any number of subsets selected by any retrieval heuristic or method, and to identify which subset provides the most benefit for the target task. In extensive experiments, COLLAGE outperforms state-of-the-art retrieval and multi-task learning approaches, achieving a 5.1% improvement over the best baseline in simulation across 10 tasks, and a 16.6% improvement in the real world across 6 tasks. For our real world experiments, we include data selection from the large-scale, real-world DROID dataset, significantly improving few-shot imitation policy training. More information at: https://collagecorl25.github.io/

In the field of robot learning, it is becoming possible to coordinate robot action through language instructions. On the other hand, it is still a difficult task to adjust the action based on human instructions because human instructions are often qualitative, and there are cases where there is no one-to-one correspondence between the behavior and the instructions. In this paper, we propose a motion generation model that can adjust actions in response to qualitative human instructions during task execution. The core of the proposed method is a learning architecture that maps qualitative human instructions to actions. Specifically, the demonstration is divided into short action sequences, and labels reflecting human qualitative senses are assigned to these sequences to realize learning that links human qualitative instructions and robot actions. In evaluation experiments, we verified the effectiveness of the method in two tasks: a pick-and-place task and a wiping task. Experimental results showed that the proposed method is able to generate motions in response to human qualitative instructions during task execution, whereas the conventional method generates trajectories all at once, making it impossible to adjust motions during task execution.

Balance control is important for human and bipedal robotic systems. While dynamic balance during locomotion has received considerable attention, quantitative understanding of static balance and falling remains limited. This work presents a hierarchical control pipeline for simulating human balance via a comprehensive whole-body musculoskeletal system. We identified spatiotemporal dynamics of balancing during stable standing, revealed the impact of muscle injury on balancing behavior, and generated fall contact patterns that aligned with clinical data. Furthermore, our simulated hip exoskeleton assistance demonstrated improvement in balance maintenance and reduced muscle effort under perturbation. This work offers unique muscle-level insights into human balance dynamics that are challenging to capture experimentally. It could provide a foundation for developing targeted interventions for individuals with balance impairments and support the advancement of humanoid robotic systems.

Recent breakthroughs in large language models (LLMs) have not only advanced natural language processing but also inspired their application in domains with structurally similar problems—most notably, autonomous driving motion generation. Both domains involve autoregressive sequence modeling, token-based representations, and context-aware decision making, making the transfer of LLM components a natural and increasingly common practice. However, despite promising early attempts, a systematic understanding of which LLM modules are truly transferable remains lacking. In this paper, we present a comprehensive evaluation of five key LLM modules—tokenizer design, positional embedding, pre-training paradigms, post-training strategies, and test-time computation—within the context of motion generation for autonomous driving. Through extensive experiments on the Waymo Sim Agents benchmark, we demonstrate that, when appropriately adapted, these modules can significantly improve performance for autonomous driving motion generation. In addition, we identify which techniques can be effectively transferred, analyze the potential reasons for the failure of others, and discuss the specific adaptations needed for autonomous driving scenarios. We evaluate our method on the Sim Agents task and achieve competitive results.