CoRL 2025 论文列表

A key open challenge in off-road autonomy is that the traversability of terrain often depends on the vehicle's state. In particular, some obstacles are only traversable from some orientations. However, learning this interaction by encoding the angle of approach as a model input demands a large and diverse training dataset and is computationally inefficient during planning due to repeated model inference. To address these challenges, we present SPARTA, a method for estimating approach angle conditioned traversability from point clouds. Specifically, we impose geometric structure into our network by outputting a smooth analytical function over the 1-Sphere that predicts risk distribution for any angle of approach with minimal overhead and can be reused for subsequent queries. The function is composed of Fourier basis functions, which has important advantages for generalization due to their periodic nature and smoothness. We demonstrate SPARTA both in a high-fidelity simulation platform, where our model achieves a 91% success rate crossing a 40m boulder field (compared to 73\% for the baseline), and on hardware, illustrating the generalization ability of the model to real-world settings.

Simulation-based reinforcement learning (RL) has significantly advanced humanoid locomotion tasks, yet direct real-world RL from scratch or starting from pretrained policies remains rare, limiting the full potential of humanoid robots. Real-world training, despite being crucial for overcoming the sim-to-real gap, faces substantial challenges related to safety, reward design, and learning efficiency. To address these limitations, we propose Robot-Trains-Robot (RTR), a novel framework where a robotic arm teacher actively supports and guides a humanoid student robot. The RTR system provides protection, schedule, reward, perturbation, failure detection, and automatic resets, enabling efficient long-term real-world training with minimal human intervention. Furthermore, we propose a novel RL pipeline that facilitates and stabilizes sim-to-real transfer by optimizing a single dynamics-encoded latent variable in the real world. We validate our method through two challenging real-world humanoid tasks: fine-tuning a walking policy for precise speed tracking and learning a humanoid swing-up task from scratch, illustrating the promising capabilities of real-world humanoid learning realized by RTR-style systems.

Training robots in simulation requires diverse 3D scenes that reflect the specific challenges of downstream tasks. However, scenes that satisfy strict task requirements, such as high-clutter environments with plausible spatial arrangement, are rare and costly to curate manually. Instead, we generate large-scale scene data using procedural models that approximate realistic environments for robotic manipulation, and adapt it to task-specific goals. We do this by training a unified diffusion-based generative model that predicts which objects to place from a fixed asset library, along with their SE(3) poses. This model serves as a flexible scene prior that can be adapted using reinforcement learning-based post training, conditional generation, or inference-time search, steering generation toward downstream objectives even when they differ from the original data distribution. Our method enables goal-directed scene synthesis that respects physical feasibility and scales across scene types. We introduce a novel MCTS-based inference-time search strategy for diffusion models, enforce feasibility via projection and simulation, and release a dataset of over 44 million SE(3) scenes spanning five diverse environments.

Today's visuo-tactile sensors come in many shapes and sizes, making it challenging to develop general-purpose tactile representations. This is because most models are tied to a specific sensor design. To address this challenge, we propose two approaches to cross-sensor image generation. The first is an end-to-end method that leverages paired data (Touch2Touch). The second method builds an intermediate depth representation and does not require paired data (T2D2: Touch-to-Depth-to-Touch). Both methods enable the use of sensor-specific models across multiple sensors via the cross-sensor touch generation process. Together, these models offer flexible solutions for sensor translation, depending on data availability and application needs. We demonstrate their effectiveness on downstream tasks such as cup stacking and tool insertion, where models originally designed for one sensor are successfully transferred to another using in-hand pose estimation.

Cloth manipulation is challenging due to its highly complex dynamics, near-infinite degrees of freedom, and frequent self-occlusions, which complicate both state estimation and dynamics modeling. Inspired by recent advances in generative models, we hypothesize that these expressive models can effectively capture intricate cloth configurations and deformation patterns from data. Therefore, we propose a diffusion-based generative approach for both perception and dynamics modeling. Specifically, we formulate state estimation as reconstructing full cloth states from partial observations and dynamics modeling as predicting future states given the current state and robot actions. Leveraging a transformer-based diffusion model, our method achieves accurate state reconstruction and reduces long-horizon dynamics prediction errors by an order of magnitude compared to prior approaches. We integrate our dynamics models with model-predictive control and show that our framework enables effective cloth folding on real robotic systems, demonstrating the potential of generative models for deformable object manipulation under partial observability and complex dynamics.

Predictive manipulation has recently gained considerable attention in the Embodied AI community due to its potential to improve robot policy performance by leveraging predicted states. However, generating accurate future visual states of robot-object interactions from world models remains a well-known challenge, particularly in achieving high-quality pixel-level representations. To this end, we propose LaDi-WM, a world model that predicts the latent space of future states using diffusion modeling. Specifically, LaDi-WM leverages the well-established latent space aligned with pre-trained Visual Foundation Models (VFMs), which comprises both geometric features (DINO-based) and semantic features (CLIP-based). We find that predicting the evolution of the latent space is easier to learn and more generalizable than directly predicting pixel-level images. Building on LaDi-WM, we design a diffusion policy that iteratively refines output actions by incorporating forecasted states, thereby generating more consistent and accurate results. Extensive experiments on both synthetic and real-world benchmarks demonstrate that LaDi-WM significantly enhances policy performance by 27.9\% on the LIBERO-LONG benchmark and 20\% on the real-world scenario. Furthermore, our world model and policies achieve impressive generalizability in real-world experiments. The full source code will be publicly available.

Reasoning over long sequences of observations and actions is essential for many robotic tasks. Yet, learning effective long-context policies from demonstrations remains challenging. As context length increases, training becomes increasingly expensive due to rising memory demands, and policy performance often degrades as a result of spurious correlations. Recent methods typically sidestep these issues by truncating context length, discarding historical information that may be critical for subsequent decisions. In this paper, we propose an alternative approach that explicitly regularizes the retention of past information. We first revisit the copycat problem in imitation learning and identify an opposite challenge in recent diffusion policies: rather than over-relying on prior actions, they often fail to capture essential dependencies between past and future actions. To address this, we introduce Past-Token Prediction (PTP), an auxiliary task in which the policy learns to predict past action tokens alongside future ones. This regularization significantly improves temporal modeling in the policy head, with minimal reliance on visual representations. Building on this observation, we further introduce a multistage training strategy: pre-train the visual encoder with short contexts, and fine-tune the policy head using cached long-context embeddings. This strategy preserves the benefits of PTP while greatly reducing memory and computational overhead. Finally, we extend PTP into a self-verification mechanism at test time, enabling the policy to score and select candidates consistent with past actions during inference. Experiments across four real-world and six simulated tasks demonstrate that our proposed method improves the performance of long-context diffusion policies by 3× and accelerates policy training by more than 10×. Videos are available at https://ptp-robot.github.io.

Diffusion models hold great potential in robotics due to their ability to capture complex, high-dimensional data distributions. However, their lack of constraint-awareness limits their deployment in safety-critical applications. We propose Constraint-Aware Diffusion Guidance (CoDiG), a data-efficient and general-purpose framework that integrates barrier functions into the denoising process, guiding diffusion sampling toward constraint-satisfying outputs. CoDiG enables constraint satisfaction even with limited training data and generalizes across tasks. We evaluate our framework in the challenging setting of miniature autonomous racing, where real-time obstacle avoidance is essential. Real-world experiments show that CoDiG generates safe outputs efficiently under dynamic conditions, highlighting its potential for broader robotic applications.

Panoramic cameras, capturing comprehensive 360-degree environmental data, are suitable for quadruped robots in surrounding perception and interaction with complex environments. However, the scarcity of high-quality panoramic training data — caused by inherent kinematic constraints and complex sensor calibration challenges — fundamentally limits the development of robust perception systems tailored to these embodied platforms. To address this issue, we propose QuaDreamer—the first panoramic data generation engine specifically designed for quadruped robots. QuaDreamer focuses on mimicking the motion paradigm of quadruped robots to generate highly controllable, realistic panoramic videos, providing a data source for downstream tasks. Specifically, to effectively capture the unique vertical vibration characteristics exhibited during quadruped locomotion, we introduce Vertical Jitter Encoding (VJE). VJE extracts controllable vertical signals through frequency-domain feature filtering and provides high-quality prompts. To facilitate high-quality panoramic video generation under jitter signal control, we propose a Scene-Object Controller (SOC) that effectively manages object motion and boosts background jitter control through the attention mechanism. To address panoramic distortions in wide-FoV video generation, we propose the Panoramic Enhancer (PE) – a dual-stream architecture that synergizes frequency-texture refinement for local detail enhancement with spatial-structure correction for global geometric consistency. We further demonstrate that the generated video sequences can serve as training data for the quadruped robot's panoramic visual perception model, enhancing the performance of multi-object tracking in 360-degree scenes. The source code and model weights will be publicly available.

When navigating in a man-made environment they haven’t visited before—like an office building—humans employ behaviors such as reading signs and asking others for directions. These behaviors help humans reach their destinations efficiently by reducing the need to search through large areas. Existing robot navigation systems lack the ability to execute such behaviors and are thus highly inefficient at navigating within large environments. We present ReasonNav, a modular navigation system which integrates these human-like navigation skills by leveraging the reasoning capabilites of a vision-language model (VLM). We design compact input and output abstractions based on navigation landmarks, allowing the VLM to focus on language understanding and reasoning. We evaluate ReasonNav on real and simulated navigation tasks and show that the agent successfully employs higher-order reasoning to navigate efficiently in large, complex buildings.

In this paper, we propose a training-free framework for vision-and-language navigation (VLN). Existing zero-shot VLN methods are mainly designed for discrete environments or involve unsupervised training in continuous simulator environments, which makes it challenging to generalize and deploy them in real-world scenarios. To achieve a training-free framework in continuous environments, our framework formulates navigation guidance as graph constraint optimization by decomposing instructions into explicit spatial constraints. The constraint-driven paradigm decodes spatial semantics through constraint solving, enabling zero-shot adaptation to unseen environments. Specifically, we construct a spatial constraint library covering all types of spatial relationship mentioned in VLN instructions. The human instruction is decomposed into a directed acyclic graph, with waypoint nodes, object nodes and edges, which are used as queries to retrieve the library to build the graph constraints. The graph constraint optimization is solved by the constraint solver to determine the positions of waypoints, obtaining the robot's navigation path and final goal. To handle cases of no solution or multiple solutions, we construct the navigation tree and the backtracking mechanism. Extensive experiments on standard benchmarks demonstrate significant improvements in success rate and navigation efficiency compared to state-of-the-art zero-shot VLN methods. We further conduct real-world experiments to show our framework can effectively generalize to new environments and instruction sets, paving the way for more robust and autonomous navigation framework.

Building robotic agents capable of operating across diverse environments and object types remains a significant challenge, often requiring extensive data collection. This is particularly restrictive in robotics, where each data point must be physically executed in the real world. Consequently, there is a critical need for alternative data sources for robotics and frameworks that enable learning from such data. In this work, we present Point Policy, a new method for learning robot policies exclusively from offline human demonstration videos without any teleoperation data. Point Policy leverages state-of-the-art vision models and policy architectures to translate human hand poses into robot poses while capturing object states through semantically meaningful key points. This approach yields a morphology-agnostic representation that facilitates effective policy learning. Through experiments on a diverse set of real-world tasks, we demonstrate that Point Policy significantly outperforms prior methods for policy learning from human videos, performing well not only within the training distribution but also generalizing to novel object instances and cluttered environments. Videos of the robot are best viewed at anon-point-policy.github.io.

Kinodynamic motion planning is concerned with computing collision-free trajectories while abiding by the robot's dynamic constraints. This critical problem is often tackled using sampling-based planners (SBPs) that explore the robot's high-dimensional state space by constructing a search tree via action propagations. Although SBPs can offer global guarantees on completeness and solution quality, their performance is often hindered by slow exploration due to uninformed action sampling. Learning-based approaches can yield significantly faster runtimes, yet they fail to generalize to out-of-distribution (OOD) scenarios and lack critical guarantees, e.g., safety, thus limiting their deployment on physical robots. We present Diffusion Tree (DiTree): a provably-generalizable framework leveraging diffusion policies (DPs) as informed samplers to efficiently guide state-space search within SBPs. DiTree combines DP's ability to model complex distributions of expert trajectories, conditioned on local observations, with the completeness of SBPs, to yield provably-safe solutions within a few action propagation iterations for complex dynamical systems. We demonstrate DiTree's power with an implementation combining the popular RRT planner with a DP action sampler trained on a single environment. In comprehensive evaluations on OOD scenarios, DiTree has comparable runtimes to a standalone DP (4x faster than classical SBPs), while improving the success rate over DP and SBPs (on average).

Estimating the 6D pose of arbitrary objects from a single reference image is a critical yet challenging task in robotics, especially considering the long-tail distribution of real-world instances. While category-level and model-based approaches have achieved notable progress, they remain limited in generalizing to unseen objects under one-shot settings. In this work, we propose a novel pipeline for fast and accurate one-shot 6D pose and scale estimation. Leveraging recent advances in single-view 3D generation, we first build high-fidelity textured meshes without requiring known object poses. To resolve scale ambiguity, we introduce a coarse-to-fine alignment module that estimates both object size and initial pose by matching 2D-3D features with depth information. We then generate a diversified set of plausible 3D models using text-guided generative augmentation and render them with Blender to synthesize large-scale, domain-randomized training data for pose estiamtion. This synthetic data bridges the domain gap and enables robust fine-tuning of pose estimators. Our method achieves state-of-the-art results on several 6D pose benchmarks, and we further validate its effectiveness on a newly collected in-the-wild dataset. Finally, we integrate our system with a dexterous hand, demonstrating its robustness in real-world robotic grasping tasks. All code, data, and models will be released to foster future research.

Scaling up robotic imitation learning for real-world applications requires efficient and scalable demonstration collection methods. While teleoperation is effective, it depends on costly and inflexible robot platforms. In-the-wild demonstrations offer a promising alternative, but existing collection devices have key limitations: handheld setups offer limited observational coverage, and whole-body systems often require fine-tuning with robot data due to domain gaps. To address these challenges, we present AirExo-2, a low-cost exoskeleton system for large-scale in-the-wild data collection, along with visual adaptors that transform collected data into pseudo-robot demonstrations suitable for policy learning. We further introduce RISE-2, a generalizable imitation learning policy that fuses 3D spatial and 2D semantic perception for robust manipulations. Experiments show that RISE-2 outperforms prior state-of-the-art methods on both in-domain and generalization evaluations. Trained solely on adapted in-the-wild data produced by AirExo-2, RISE-2 achieves comparable performance to policies trained with teleoperated data, highlighting the effectiveness and potential of AirExo-2 for scalable and generalizable imitation learning.

Imagine a robot learning locomotion skills from any single video, without labels or reward engineering. We introduce SDS ("See it. Do it. Sorted."), an automated pipeline for skill acquisition from unstructured video demonstrations. Using GPT-4o, SDS applies novel prompting techniques, in the form of spatio-temporal grid-based visual encoding (Gv) and structured input decomposition (SUS). These produce executable reward functions (RF) from raw input videos. The RFs are used to train PPO policies and are optimized through closed-loop evolution, using training footage and performance metrics as self-supervised signals. SDS allows quadrupeds (e.g., Unitree Go1) to learn four gaits—trot, bound, pace, and hop—achieving 100% gait matching fidelity, Dynamic Time Warping (DTW) distance in the order of 10^-6, and stable locomotion with zero failures, both in simulation and the real world. SDS generalizes to morphologically different quadrupeds (e.g., ANYmal) and outperforms prior work in data efficiency, training time, and engineering effort. Our code is open-source under: https://sdsreview.github.io/SDS_ANONYM/

Learning real-world robotic manipulation is challenging, particularly when limited demonstrations are available. Existing methods for few-shot manipulation often rely on simulation-augmented data or pre-built modules like grasping and pose estimation, which struggle with sim-to-real gaps and lack extensibility. While large-scale imitation pre-training shows promise, adapting these general-purpose policies to specific tasks in data-scarce settings remains unexplored. To achieve this, we propose ControlVLA, a novel framework that bridges pre-trained VLA models with object-centric representations via a ControlNet-style architecture for efficient fine-tuning. Specifically, to introduce object-centric conditions without overwriting prior knowledge, ControlVLA zero-initializes a set of projection layers, allowing them to gradually adapt the pre-trained manipulation policies. In real-world experiments across 6 diverse tasks, including pouring cubes and folding clothes, our method achieves a 76.7\% success rate while requiring only 10-20 demonstrations --- a significant improvement over traditional approaches that require more than 100 demonstrations to achieve comparable success. Additional experiments highlight ControlVLA's extensibility to long-horizon tasks and robustness to unseen objects and backgrounds.

Efficient robot locomotion often requires balancing task performance with energy expenditure. A common approach in reinforcement learning (RL) is to penalize energy use directly in the reward function. This requires carefully weighting the reward terms to avoid undesirable trade-offs where energy minimization harms task success or vice versa. In this work, we propose a hyperparameter-free gradient optimization method to minimize energy without conflicting with task performance. Inspired by recent works in multitask learning, our method applies policy gradient projection between task and energy objectives to promote non-conflicting updates. We evaluate this technique on standard locomotion benchmarks of DM-Control and HumanoidBench and demonstrate a reduction of $64$% energy usage while maintaining comparable task performance. Further, we conduct experiments on a Unitree GO2 quadruped showcasing Sim2Real transfer of energy efficient policies. Our method is easy to implement in standard RL pipelines with minimal code changes, and offers a principled alternative to reward shaping for energy efficient control policies.

Data-driven control methods need to be sample-efficient and lightweight, especially when data acquisition and computational resources are limited---such as during learning on hardware. Most modern data-driven methods require large datasets and struggle with real-time updates of models, limiting their performance in dynamic environments. Koopman theory formally represents nonlinear systems as linear models over observables, and Koopman representations can be determined from data in an optimization-friendly setting with potentially rapid model updates. In this paper, we present a highly sample-efficient, Koopman-based learning pipeline: Recursive Koopman Learning (RKL). We identify sufficient conditions for model convergence and provide formal algorithmic analysis supporting our claim that RKL is lightweight and fast, with complexity independent of dataset size. We validate our method on a simulated planar two-link arm and a hybrid nonlinear hardware system with soft actuators, showing that real-time recursive Koopman model updates improve the sample efficiency and stability of data-driven controller synthesis---requiring only <10% of the data compared to benchmarks. The high-performance C++ codebase will be open-sourced.