CoRL 2025 论文列表

We present See, Point, Fly (SPF), a training-free aerial vision-and-language navigation (AVLN) framework built atop vision-language models (VLMs). SPF is capable of navigating to any goal based on any type of free-form instructions in any kind of environment. In contrast to existing VLM-based approaches that treat action prediction as a text generation task, our key insight is to consider action prediction for AVLN as a 2D spatial grounding task. SPF harness VLMs to decompose vague language instructions into iterative annotation of 2D waypoints on the input image. Along with the predicted traveling distance, SPF transforms predicted 2D waypoints into 3D displacement vectors as action commands for UAVs. Moreover, SPF also adaptively adjusts the traveling distance to facilitate more efficient navigation. Notably, SPF performs navigation in a closed-loop control manner, enabling UAVs to follow dynamic targets in dynamic environments. SPF sets a new state of the art in DRL simulation benchmark, out performing the previous best method by an absolute margin of 63%. In extensive real-world evaluations, SPF outperforms strong baselines by a large margin. We also conduct comprehensive ablation studies to highlight the effectiveness of our design choice. Lastly, SPF shows remarkable generalization to different VLMs.

Existing evaluation paradigms for Autonomous Vehicles (AVs) face critical limitations. Real-world evaluation is often challenging due to safety concerns and a lack of reproducibility, whereas closed-loop simulation can face insufficient realism or high computational costs. Open-loop evaluation, while being efficient and data-driven, relies on metrics that generally overlook compounding errors. In this paper, we propose pseudo-simulation, a novel paradigm that addresses these limitations. Pseudo-simulation operates on real datasets, similar to open-loop evaluation, but augments them with synthetic observations generated prior to evaluation using 3D Gaussian Splatting. Our key idea is to approximate potential future states the AV might encounter by generating a diverse set of observations that vary in position, heading, and speed. Our method then assigns a higher importance to synthetic observations that best match the AV's likely behavior using a novel proximity-based weighting scheme. This enables evaluating error recovery and the mitigation of causal confusion, as in closed-loop benchmarks, without requiring sequential interactive simulation. We show that pseudo-simulation is better correlated with closed-loop simulations ($R^2=0.8$) than the best existing open-loop approach ($R^2=0.7$). We also establish a public leaderboard for the community to benchmark new methodologies with pseudo-simulation.

Achieving coordinated teamwork among legged robots requires both fine-grained locomotion control and long-horizon strategic decision-making. Robot soccer offers a compelling testbed for this challenge, combining dynamic, competitive, and multi-agent interactions. In this work, we present a hierarchical multi-agent reinforcement learning (MARL) framework that enables fully autonomous and decentralized quadruped robot soccer. First, a set of highly dynamic low-level skills is trained for legged locomotion and ball manipulation, such as walking, dribbling, and kicking. On top of these, a high-level strategic planning policy is trained with Multi-Agent Proximal Policy Optimization (MAPPO) via Fictitious Self-Play (FSP). This learning framework allows agents to adapt to diverse opponent strategies and gives rise to sophisticated team behaviors, including coordinated passing, interception, and dynamic role allocation. With an extensive ablation study, the proposed learning method shows significant advantages in the cooperative and competitive multi-agent soccer game. We deploy the learned policies to real quadruped robots relying solely on onboard proprioception and decentralized localization, with the resulting system supporting autonomous robot-robot and robot-human soccer matches on indoor and outdoor soccer courts.

Recent advances in Behavior Cloning (BC) have led to strong performance in robotic manipulation, driven by expressive models, sequence modeling of actions, and large-scale demonstration data. However, BC faces significant challenges when applied to heterogeneous datasets, such as visual shift with different camera poses or object appearances, where performance degrades despite the benefits of learning at scale. This stems from BC's tendency to overfit individual demonstrations rather than capture shared structure, limiting generalization. To address this, we introduce Contrastive Learning via Action Sequence Supervision (CLASS), a method for learning behavioral representations from demonstrations using supervised contrastive learning. CLASS leverages weak supervision from similar action sequences identified via Dynamic Time Warping (DTW) and optimizes a soft InfoNCE loss with similarity-weighted positive pairs. We evaluate CLASS on 5 simulation benchmarks and 3 real-world tasks to achieve competitive results using retrieval-based control with representations only. Most notably, for downstream policy learning under significant visual shifts, CLASS achieves an average success rate of 70% with Diffusion Policy, while all other baseline methods fail to perform competitively.

Large vision and language models are being increasingly deployed on real robots, leading to an immediate need for ensuring robot safety under AI-control. In this paper, we develop the ASIMOV Benchmark — a collection of large-scale semantic safety datasets grounded in real-world visual scenes and human injury reports from hospitals (500k situations, 3M instructions). We propose a scalable recipe for data generation leveraging text and image generation techniques to synthesize safety-relevant scenarios. As a second contribution, we develop a framework to automatically generate robot constitutions from real-world data to steer a robot’s behavior using Constitutional AI mechanisms. We report a top alignment rate of 84.3% on the ASIMOV Benchmark using generated constitutions, outperforming no-constitution baselines and human-written constitutions. We argue that human interpretability and modifiability of constitutions inferred from data make them an ideal medium for behavior governance of AI-controlled robots.

Accurate LiDAR-camera calibration is the foundation of accurate multimodal fusion environmental perception for autonomous driving and robotic systems. Traditional calibration methods require extensive data collection in controlled environments and cannot compensate for the transformation changes during the vehicle/robot movement. In this paper, we propose the first model that uses bird's-eye view (BEV) features to perform LiDAR camera calibration from raw data, termed BEVCalib. To achieve this, we extract camera BEV features and LiDAR BEV features separately and fuse them into a shared BEV feature space. To fully utilize the geometry information from the BEV feature, we introduce a novel feature selector to choose the most important feature in the transformation decoder, which reduces memory consumption and enables efficient training. Extensive evaluations in various datasets demonstrate that BEVCalib establishes a new state-of-the-art; improving the best open-source baseline by two orders of magnitude on KITTI, Nuscenes, and our dynamic extrinsic dataset, respectively, and outperforming the best baseline in literature by 72% on KITTI dataset, and 69% on Nuscenes dataset. All source code and checkpoints will be released.

Exploration is key for general-purpose robotic learning, particularly in open-ended environments where explicit guidance or task-specific feedback is limited. Vision-language models (VLMs), which can reason about object semantics, spatial relations, and potential outcomes, offer a promising foundation for guiding exploratory behavior by generating high-level goals or transitions. However, their outputs lack grounding, making it difficult to determine whether imagined transitions are physically feasible or informative. To bridge the gap between imagination and execution, we present IVE (Imagine, Verify, Execute), an agentic exploration framework inspired by human curiosity. Human exploration often emerges from the drive to discover novel scene configurations and to understand the environment. Inspired by this, IVE leverages VLMs to abstract RGB-D observations into semantic scene graphs, imagine novel scenes, predict their physical plausibility, and generate executable skill sequences through action tools. We evaluate IVE in both simulated and real-world tabletop environments. The results show that IVE produces more diverse and meaningful exploration than RL baselines. The collected data facilitates learning downstream tasks that closely match those of policies trained on human-collected demonstrations.

Large language models (LLMs) provide robots with powerful contextual reasoning abilities and a natural human interface. Yet, current LLM-enabled robots typically depend on cloud-hosted models, limiting their usability in environments with unreliable communication infrastructure, such as outdoor or industrial settings. We present PRISM, a framework for distilling small language model (SLM)-enabled robot planners that run on-device with minimal human supervision. Starting from an existing LLM-enabled planner, PRISM automatically synthesizes diverse tasks and environments, elicits plans from the LLM, and uses this synthetic dataset to distill a compact SLM as a drop-in replacement of the source model. We apply PRISM to three LLM-enabled planners for mapping and exploration, manipulation, and household assistance, and we demonstrate that PRISM improves the performance of Llama-3.2-3B from 10-20\% of GPT-4o's performance to over 93% - using only synthetic data. We further demonstrate that the distilled planners generalize across heterogeneous robotic platforms (ground and aerial) and diverse environments (indoor and outdoor). We release all software, trained models, and datasets to promote reproducibility and follow-up work.

Model Predictive Control (MPC) is among the most widely adopted and reliable methods for robot control, relying critically on an accurate dynamics model. However, existing dynamics models used in the gradient-based MPC are limited by computational complexity and state representation. To address this limitation, we propose the Hyper Prediction Model (HyperPM) - a novel approach in which we project the unmodeled dynamics onto a time-dependent dynamics model. This time-dependency is captured through time-varying model parameters, whose evolution over the MPC prediction horizon is learned using a neural network. Such formulation preserves the computational efficiency and robustness of the base model while equipping it with the capacity to anticipate previously unmodeled phenomena. We evaluated the proposed approach on several challenging systems, including real-world F1TENTH autonomous racing, and demonstrated that it significantly reduces long-horizon prediction errors. Moreover, when integrated within the MPC framework (HyperMPC), our method consistently outperforms existing state-of-the-art techniques.

Nominal payload ratings for articulated robots are typically derived from worst-case configurations, resulting in uniform payload constraints across the entire workspace. This conservative approach severely underutilizes the robot's inherent capabilities---our analysis demonstrates that manipulators can safely handle payloads well above nominal capacity across broad regions of their workspace while staying within joint angle, velocity, acceleration, and torque limits. To address this gap between assumed and actual capability, we propose a novel trajectory generation approach using denoising diffusion models that explicitly incorporates payload constraints into the planning process. Unlike traditional sampling-based methods that rely on inefficient trial-and-error, optimization-based methods that are prohibitively slow, or kinodynamic planners that struggle with problem dimensionality, our approach generates dynamically feasible joint-space trajectories in constant time that can be directly executed on physical hardware without post-processing. Experimental validation on a 7 DoF Franka Emika Panda robot demonstrates that up to 67.6% of the workspace remains accessible even with payloads exceeding 3 times the nominal capacity. This expanded operational envelope highlights the importance of a more nuanced consideration of payload dynamics in motion planning algorithms.

Learning generalizable robot manipulation policies, especially for complex multi-fingered humanoids, remains a significant challenge. Existing approaches primarily rely on extensive data collection and imitation learning, which are expensive, labor-intensive, and difficult to scale. Sim-to-real reinforcement learning (RL) offers a promising alternative, but has mostly succeeded in simpler state-based or single-hand setups. How to effectively extend this to vision-based, contact-rich bimanual manipulation tasks remains an open question. In this paper, we introduce a practical sim-to-real RL recipe that trains a humanoid robot to perform three challenging dexterous manipulation tasks: grasp-and-reach, box lift and bimanual handover. Our method features an automated real-to-sim tuning module, a generalized reward formulation based on contact and object goals, a divide-and-conquer policy distillation framework, and a hybrid object representation strategy with modality-specific augmentation. We demonstrate high success rates on unseen objects and robust, adaptive policy behaviors -- highlighting that vision-based dexterous manipulation via sim-to-real RL is not only viable, but also scalable and broadly applicable to real-world humanoid manipulation tasks.

3D world models (i.e., learning-based 3D dynamics models) offer a promising approach to generalizable robotic manipulation by capturing the underlying physics of environment evolution conditioned on robot actions. However, existing 3D world models are primarily limited to single-material dynamics using a particle-based Graph Neural Network model, and often require time-consuming 3D scene reconstruction to obtain 3D particle tracks for training. In this work, we present ParticleFormer, a Transformer-based point cloud world model trained with a hybrid point cloud reconstruction loss, supervising both global and local dynamics features in multi-material, multi-object robot interactions. ParticleFormer captures fine-grained multi-object interactions between rigid, deformable, and flexible materials, trained directly from real-world robot perception data without an elaborate scene reconstruction. We demonstrate the model's effectiveness both in 3D scene forecasting tasks, and in downstream manipulation tasks using a Model Predictive Control (MPC) policy. In addition, we extend existing dynamics learning benchmarks to include diverse multi-material, multi-object interaction scenarios. We validate our method on six simulation and three real-world experiments, where it consistently outperforms leading baselines by achieving superior dynamics prediction accuracy and less rollout error in downstream visuomotor tasks. Experimental videos are available at https://particleformer.github.io/.

Learning robot manipulation from abundant human videos offers a scalable alternative to costly robot-specific data collection. However, domain gaps across visual, morphological, and physical aspects hinder direct imitation. To effectively bridge the domain gap, we propose ImMimic, an embodiment-agnostic co-training framework that leverages both human videos and a small amount of teleoperated robot demonstrations. ImMimic uses Dynamic Time Warping (DTW) with either action- or visual-based mapping to map retargeted human hand poses to robot joints, followed by MixUp interpolation between paired human and robot trajectories. Our key insights are (1) retargeted human hand trajectories provide informative action labels, and (2) interpolation over the mapped data creates intermediate domains that facilitate smooth domain adaptation during co-training. Evaluations on four real-world manipulation tasks (Pick and Place, Push, Hammer, Flip) across four robotic embodiments (Robotiq, Fin Ray, Allegro, Ability) show that ImMimic improves task success rates and execution smoothness, highlighting its efficacy to bridge the domain gap for robust robot manipulation. The project website can be found at https://sites.google.com/view/immimic.

In endoscopic procedures, autonomous tracking of abnormal regions and following of circumferential cutting markers can significantly reduce the cognitive burden on endoscopists. However, conventional model-based pipelines are fragile—each component (e.g., detection, motion planning) requires manual tuning and struggles to incorporate high-level endoscopic intent, resulting in poor generalization across variable scenes. Vision–Language–Action (VLA) models, which integrate visual perception, language grounding, and motion planning within an end-to-end framework, offer a promising alternative to semantically adapt to surgeon prompts, without the need for manual recalibration. Despite their potential, applying VLA models to robotic endoscopy presents unique challenges due to the inherently complex and dynamic anatomical environments of the gastrointestinal (GI) tract. To this end, we introduce EndoVLA, designed specifically for continuum robots in GI interventions. Provided endoscopic images and surgeon-issued tracking prompts, EndoVLA performs three core tasks: (1) polyp tracking, (2) delineation and following of abnormal mucosal regions, and (3) adherence to predefined circular markers during circumferential cutting. To address the unique challenges posed by data scarcity and domain shifts, we propose a dual-phase strategy, with supervised fine-tuning on our EndoVLA-Motion dataset and reinforcement fine-tuning using task-aware rewards. Our approach significantly enhances the tracking performance in endoscopy, and zero-shot generalization of tracking in general scenes and more challenging sequential tasks.

Dexterous teleoperation plays a crucial role in robotic manipulation for real-world data collection and remote robot control. Previous dexterous teleoperation mostly relies on hand retargeting to closely mimic human hand postures. However, these approaches may fail to fully leverage the inherent dexterity of dexterous hands, which can execute unique actions through their structural advantages compared to human hands. To address this limitation, we propose TypeTele, a type-guided dexterous teleoperation system, which enables dexterous hands to perform actions that are not constrained by human motion patterns. This is achieved by introducing dexterous manipulation types into the teleoperation system, allowing operators to employ appropriate types to complete specific tasks. To support this system, we build an extensible dexterous manipulation type library to cover comprehensive dexterous postures used in manipulation tasks. During teleoperation, we employ a MLLM (Multi-modality Large Language Model)-assisted type retrieval module to identify the most suitable manipulation type based on the specific task and operator commands. Extensive experiments of real-world teleoperation and imitation learning demonstrate that the incorporation of manipulation types significantly takes full advantage of the dexterous robot's ability to perform diverse and complex tasks with higher success rates.

Developing robotic systems capable of robustly executing long-horizon manipulation tasks with human-level dexterity is challenging, as such tasks require both physical dexterity and seamless sequencing of manipulation skills while robustly handling environment variations. While imitation learning offers a promising approach, acquiring comprehensive datasets is resource-intensive. In this work, we propose a learning framework and system LodeStar that automatically decomposes task demonstrations into semantically meaningful skills using off-the-shelf foundation models, and generates diverse synthetic demonstration datasets from a few human demos through reinforcement learning. These sim-augmented datasets enable robust skill training, with a Skill Routing Transformer (SRT) policy effectively chaining the learned skills together to execute complex long-horizon manipulation tasks. Experimental evaluations on three challenging real-world long-horizon dexterous manipulation tasks demonstrate that our approach significantly improves task performance and robustness compared to previous baselines. Videos are available at lodestar-robot.github.io.

Multi-task ``vision-language-action'' (VLA) models have recently demonstrated increasing promise as generalist foundation models for robotics, achieving non-trivial performance out of the box on new tasks in new environments. However, for such models to be truly useful, an end user must have easy means to teach them to improve. For language and vision models, the emergent ability to perform in-context learning (ICL) has proven to be a versatile and highly useful interface to easily teach new tasks with no parameter finetuning. Unfortunately, VLAs pre-trained with imitation learning objectives do not naturally acquire ICL abilities. In this paper, we demonstrate that, with the right finetuning recipe and a small robot demonstration dataset, it is possible to inject in-context adaptability post hoc into such a VLA. After retraining for in-context learning (RICL), our system permits an end user to provide a small number (10-20) of demonstrations for a new task. RICL then fetches the most relevant portions of those demonstrations into the VLA context to exploit ICL, performing the new task and boosting task performance. We apply RICL to inject ICL into the $\pi_0$-FAST VLA, and show that it permits large in-context improvements for a variety of new manipulation tasks with only 20 demonstrations per task, without any parameter updates. When parameter updates on the target task demonstrations is possible, RICL finetuning further boosts performance. We release code and model weights for RICL-$\pi_0$-FAST alongside the paper to enable, for the first time, a simple in-context learning interface for new manipulation tasks

Generalizable object fetching in cluttered scenes remains a fundamental and application-critical challenge in embodied AI. Closely packed objects cause inevitable occlusions, making safe action generation particularly difficult. Under such partial observability, effective policies must not only generalize across diverse objects and layouts but also reason about occlusion to avoid collisions. However, collecting large-scale real-world data for this task remains prohibitively expensive, leaving this problem largely unsolved. In this paper, we introduce FetchBot, a sim-to-real framework for this challenge. We first curate a large-scale synthetic dataset featuring 1M diverse scenes and 500k representative demonstrations. Based on this dataset, FetchBot employs a depth-conditioned method for action generation, which leverages structural cues to enable robust obstacle-aware action planning. However, depth is perfect in simulation but noisy in real-world environments. To address this sim-to-real gap, FetchBot predicts depth from RGB inputs using a foundation model and integrates local occupancy prediction as a co-training task, providing a generalizable latent representation for sim-to-real transfer. Extensive experiments in simulation and real-world environments demonstrate FetchBot’s strong zero-shot sim-to-real transfer, effective clutter handling, and adaptability to novel scenarios. In cluttered environments, it achieves an average success rate of 89.95%, significantly outperforming prior methods. Moreover, FetchBot demonstrates excellent robustness in challenging cases, such as fetching transparent, reflective, and irregular objects, highlighting its practical value.

Mastering deformable object manipulation often necessitates the use of anthropomorphic, high-degree-of-freedom robot hands capable of precise, contact-rich control. However, current trajectory optimisation methods often struggle in these settings due to the large search space and the sparse task information available from shape-matching cost functions, particularly when contact is absent. In this work, we propose D-Cubed, a novel trajectory optimisation method using a latent diffusion model (LDM) trained from a task-agnostic play dataset to solve dexterous deformable object manipulation tasks. D-Cubed learns a skill-latent space that encodes short-horizon actions from a play dataset using a VAE and trains a LDM to compose the skill latents into a skill trajectory, representing a long-horizon action trajectory. To optimise a trajectory for a target task, we introduce a novel gradient-free guided sampling method that employs the Cross-Entropy method within the reverse diffusion process. In particular, D-Cubed samples a small number of noisy skill trajectories using the LDM for exploration and evaluates the trajectories in simulation. Then D-Cubed selects the trajectory with the lowest cost for the subsequent reverse process. This effectively explores promising solution areas and optimises the sampled trajectories towards a target task throughout the reverse diffusion process. Through empirical evaluation on a published benchmark of dexterous deformable object manipulation tasks, we demonstrate that D-Cubed outperforms traditional trajectory optimisation and competitive baseline approaches by a significant margin.