CoRL 2025 Papers

In this paper, we introduce HALO, a novel Offline Reward Learning algorithm that quantifies human intuition in navigation into a vision-based reward function for robot navigation. HALO learns a reward model from offline data, leveraging expert trajectories collected from mobile robots. During training, actions are randomly sampled from the action space around the expert action and ranked using a Boltzmann probability distribution that combines their distance to the expert action with human preference scores derived from intuitive navigation queries based on the corresponding egocentric camera feed. These scores establish preference rankings, enabling the training of a novel reward model based on Plackett-Luce loss, which allows for preference-driven navigation. To demonstrate the effectiveness of HALO, we deploy its reward model in two downstream applications: (i) an offline learned policy trained directly on the HALO-derived rewards, and (ii) a model-predictive-control (MPC) based planner that incorporates the HALO reward as an additional cost term. This showcases the versatility of HALO across both learning-based and classical navigation frameworks. Our real-world deployments on a Clearpath Husky across multiple scenarios demonstrate that policies trained with HALO achieve improved performance over state-of-the-art methods in terms of success rate and normalized trajectory length while maintaining lower Fréchet distance with the human expert trajectories.

In complex environments, adaptive and fault-tolerant capabilities are essential for legged robot locomotion. To address this challenge, this study proposes a reinforcement learning framework that integrates contrastive learning with forward prediction to achieve fault-tolerant locomotion for legged robots. This framework constructs a forward prediction model with contrastive learning, incorporating a comparator and a forward model. The forward model predicts the robot's subsequent state, and the comparator compares these predictions with actual states to generate critical prediction errors. These errors are systematically integrated into the controller, facilitating the continuous adjustment and refinement of control signals.Experiments on quadruped robots across different terrains and various joint damage scenarios have verified the effectiveness of our method, especially the functions of the comparator and the forward model. Furthermore, robots can adapt to locked joints without prior training, demonstrating zero-shot transfer capability. Finally, the proposed method demonstrates universal applicability to both quadruped and hexapod robots, highlighting its potential for broader applications in legged robotics.

We introduce a model for monocular RGB relative pose estimation of a ground robot that trains from scratch without pose labels nor prior knowledge about the robot's shape or appearance. At training time, we assume: (i) a robot fitted with multiple LEDs, whose states are independent and known at each frame; (ii) knowledge of the approximate viewing direction of each LED; and (iii) availability of a calibration image with a known target distance, to address the ambiguity of monocular depth estimation. Training data is collected by a pair of robots moving randomly without needing external infrastructure or human supervision. Our model trains on the task of predicting from an image the state of each LED on the robot. In doing so, it learns to predict the position of the robot in the image, its distance, and its relative bearing. At inference time, the state of the LEDs is unknown, can be arbitrary, and does not affect the pose estimation performance. Quantitative experiments indicate that our approach: is competitive with SoA approaches that require supervision from pose labels or a CAD model of the robot; generalizes to different domains; and handles multi-robot pose estimation.

LiDAR Place Recognition (LPR) is a key component in robotic localization, enabling robots to align current scans with prior maps of their environment. While Visual Place Recognition (VPR) has embraced Vision Foundation Models (VFMs) to enhance descriptor robustness, LPR has relied on task-specific models with limited use of pre-trained foundation-level knowledge. This is due to the lack of 3D foundation models and the challenges of using VFM with LiDAR point clouds. To tackle this, we introduce ImLPR, a novel pipeline that employs a pre-trained DINOv2 VFM to generate rich descriptors for LPR. To our knowledge, ImLPR is the first method to leverage a VFM to support LPR. ImLPR converts raw point clouds into Range Image Views (RIV) to leverage VFM in the LiDAR domain. It employs MultiConv adapters and Patch-InfoNCE loss for effective feature learning. We validate ImLPR using public datasets where it outperforms state-of-the-art (SOTA) methods in intra-session and inter-session LPR with top Recall@1 and F1 scores across various LiDARs. We also demonstrate that RIV outperforms Bird’s-Eye-View (BEV) as a representation choice for adapting LiDAR for VFM. We release ImLPR as open source for the robotics community.

Exploration is one of the important bottlenecks for efficient learning in reinforcement learning, especially in the presence of sparse rewards. One way to traverse the environment faster is by passing through junctions, or metaphorical doors, in the state space. We propose a novel heuristic, $Door(s)$, focused on such narrow passages that serve as pathways to a large number of other states. Our approach works by estimating the state occupancy distribution and allows computation of its entropy, which forms the basis for our measure. Its computation is more sample-efficient compared to other similar methods and robustly works over longer horizons. Our results highlight the detection of dead-end states, show increased exploration efficiency, and demonstrate that $Door(s)$ encodes specific behaviors useful for downstream learning of various robotic manipulation tasks.

Diffusion and flow matching policies have recently demonstrated remarkable performance in robotic applications by accurately capturing multimodal robot trajectory distributions. However, their computationally expensive inference, due to the numerical integration of an ODE or SDE, limits their applicability as real-time controllers for robots. We introduce a methodology that utilizes conditional Optimal Transport couplings between noise and samples to enforce straight solutions in the flow ODE for robot action generation tasks. We show that naively coupling noise and samples fails in conditional tasks and propose incorporating condition variables into the coupling process to improve few-step performance. The proposed few-step policy achieves a 4\% higher success rate with a 10$\times$ speed-up compared to Diffusion Policy on a diverse set of simulation tasks. Moreover, it produces high-quality and diverse action trajectories within 1-2 steps on a set of real-world robot tasks. Our method also retains the same training complexity as Diffusion Policy and vanilla Flow Matching, in contrast to distillation-based approaches.

Fine-tuning diffusion policies with reinforcement learning (RL) presents significant challenges. The long denoising sequence for each action prediction impedes effective reward propagation. Additionally, standard RL methods require millions of physical interaction steps, making fine-tuning even more challenging. Prior work models the denoising steps in diffusion policies as a Markov Decision Process to adapt to RL policy updates, but its heavy reliance on environment interactions still leads to inefficiency. To bridge this gap, we introduce DiWA, a novel framework that leverages a world model for fine-tuning diffusion-based robotic skills entirely offline with reinforcement learning. Unlike model-free approaches that require millions of environment interactions to fine-tune a repertoire of robot skills, DiWA achieves effective adaptation using a world model trained once on a few hundred thousand offline play interactions. This results in dramatically improved sample efficiency, making the approach significantly more practical and safer for real-world robot learning. On the challenging CALVIN benchmark, DiWA improves performance across eight tasks using only offline adaptation, while requiring orders of magnitude fewer physical interactions than model-free baselines. To our knowledge, this is the first demonstration of fine-tuning diffusion policies for real-world robotic skills using an offline world model. We make the code publicly available at _redacted-for-review_.

Robotic loco-manipulation tasks often involve contact-rich interactions with the environment, requiring the joint modeling of contact force and robot position. However, recent visuomotor policies often focus solely on position or force control, overlooking their integration. In this work, we propose a unified policy for legged robots that jointly models force and position control learned without reliance on force sensors. By simulating diverse combinations of active position and force commands alongside external disturbances force, we use reinforcement learning to learn a policy that estimates forces from the robot's historical states and compensates for them through position and velocity adjustments. Such a policy enables a wide range of manipulation behaviors under varying combinations of force and position inputs, including position tracking, force application, force tracking, and compliant robot behaviors. Additionally, we demonstrate that the learned policy enhances trajectory-based imitation learning pipelines by incorporating essential contact information through its force estimation module, achieving approximately ~39.5% higher success rates across four challenging contact-rich manipulation tasks compared to position-control policies. Extensive experiments on both a quadrupedal mobile manipulation platform and a humanoid validate the versatility and robustness of the proposed policy across diverse scenarios.

We present SimShear: a sim-to-real pipeline for tactile control that allows use of shear information without explicitly modeling shear dynamics in simulation. Shear, which arises from lateral movements across contact surfaces, are critical for tasks involving dynamic object interactions but are challenging to simulate. We introduce shPix2pix: a shear-conditioned U-Net GAN that transforms simulated tactile images absent of shear plus a vector encoding shear information into realistic equivalents that include shear deformations, and show this outperforms baseline pix2pix methods for simulating tactile images and pose/shear prediction. This is applied to two control tasks using a pair of low-cost desktop robotic arms equipped with a vision-based tactile sensor: first, a tactile tracking task, where a follower arm tracks a surface moved by the leader arm; second, a collaborative co-lift task, where both arms jointly hold an object while the leader arm moves along a prescribed trajectory. Our method maintain contact errors within 1-2 mm across varied trajectories where shear sensing is essential for task performance. This work validates the use of sim-to-real shear modeling with rigid-body simulators, opening new possibilities for simulation in tactile robotics.

Vision-Language Models (VLMs) have revolutionized artificial intelligence and robotics due to their commonsense reasoning capabilities. In robotic manipulation, VLMs are used primarily as high-level planners, but recent work has also studied their lower-level reasoning ability, which refers to making decisions about precise robot movements. However, the community currently lacks a clear and common benchmark that can evaluate how well VLMs can aid low-level reasoning in robotics. Consequently, we propose a novel benchmark, ManipBench, to evaluate the low-level robot manipulation reasoning capabilities of VLMs across various dimensions, including how well they understand object-object interactions and deformable object manipulation. We extensively test 35 common and state-of-the-art VLM families on our benchmark, including variants to test different model sizes. The performance of VLMs significantly varies across tasks, and there is a strong correlation between this performance and trends in our real-world manipulation tasks. It also shows that there remains a significant gap between these models and human-level understanding.

In multi-robot systems, fleets are often deployed to collect data that improves the performance of machine learning models for downstream perception and planning. However, real-world robotic deployments generate vast amounts of data across diverse conditions, while only a small portion can be transmitted or labeled due to limited bandwidth, constrained onboard storage, and high annotation costs. To address these challenges, we propose Distributed Upload and Active Labeling (DUAL), a decentralized, two-stage data collection framework for resource-constrained robotic fleets. In the first stage, each robot independently selects a subset of its local observations to upload under storage and communication constraints. In the second stage, the cloud selects a subset of uploaded data to label, subject to a global annotation budget. We evaluate DUAL on classification tasks spanning multiple sensing modalities, as well as on RoadNet—a real-world dataset we collected from vehicle-mounted cameras for time and weather classification. We further validate our approach in a physical experiment using a Franka Emika Panda robot arm, where it learns to move a red cube to a green bowl. Finally, we test DUAL on trajectory prediction using the nuScenes autonomous driving dataset to assess generalization to complex prediction tasks. Across all settings, DUAL consistently outperforms state-of-the-art baselines, achieving up to 31.1\% gain in classification accuracy and a 13\% improvement in real-world robotics task completion rates.

Developing generalist agents that operate across diverse tasks, environments, and robot embodiments is a grand challenge in robotics and artificial intelligence. While substantial progress has been made in cross-task and cross-environment generalization, achieving broad generalization to novel embodiments remains elusive. In this work, we study embodiment scaling laws — the hypothesis that increasing the quantity of training embodiments improves generalization to unseen ones. To explore this, we procedurally generate a dataset of $\sim$1,000 varied robot embodiments, spanning humanoids, quadrupeds, and hexapods, and train embodiment-specific reinforcement learning experts for legged locomotion. We then distill these experts into a single generalist policy capable of handling diverse observation and action spaces. Our large-scale study reveals that generalization performance improves with the number of training embodiments. Notably, a policy trained on the full dataset zero-shot transfers to diverse unseen embodiments in both simulation and real-world evaluations. These results provide preliminary empirical evidence for embodiment scaling laws and suggest that scaling up embodiment quantity may serve as a foundation for building generalist robot agents.

Learned visuomotor policies are capable of performing increasingly complex manipulation tasks. However, most of these policies are trained on data collected from limited robot positions and camera viewpoints. This leads to poor generalization to novel robot positions, which limits the use of these policies on mobile platforms, especially for precise tasks like pressing buttons or turning faucets. In this work, we formulate the "policy mobilization" problem: find a mobile robot base pose in a novel environment that is in distribution with respect to a manipulation policy trained on a limited set of camera viewpoints. Compared to retraining the policy itself to be more robust to unseen robot base pose initializations, policy mobilization decouples navigation from manipulation and thus does not require additional demonstrations. With that, our formulation is still compatible with any approach that improves manipulation policy robustness. To study policy mobilization, we introduce the Mobi-$\pi$ framework, which includes: (1) metrics that quantify the difficulty of mobilizing a given policy, (2) a suite of simulated mobile manipulation tasks based on RoboCasa to evaluate policy mobilization, (3) visualization tools for analysis, and (4) several baseline methods. We also propose a novel approach that bridges navigation and manipulation by optimizing the robot's base pose to align with an in-distribution base pose for a learned policy. Our approach utilizes a 3D Gaussian Splatting model for novel viewpoint synthesis, a score function to evaluate pose suitability, as well as sampling-based optimization to identify optimal robot poses. We show that our approach on average outperforms the best baseline by 7.65$\times$ in simulation and 2.38$\times$ in the real world, demonstrating its effectiveness for policy mobilization.

Robotic chemists promise to both liberate human experts from repetitive tasks and accelerate scientific discovery, yet remain in their infancy. Chemical experiments involve long-horizon procedures over hazardous and deformable substances, where success requires not only task completion but also strict compliance with experimental norms. To address these challenges, we propose RoboChemist, a dual-loop framework that integrates Vision-Language Models (VLMs) with Vision-Language-Action (VLA) models. Unlike prior VLM-based systems (e.g., VoxPoser, ReKep) that rely on depth perception and struggle with transparent labware, and existing VLA systems (e.g., RDT, $\pi_0$) that lack semantic-level feedback for complex tasks, our method leverages a VLM to serve as (1) a planner to decompose tasks into primitive actions, (2) a visual prompt generator to guide VLA models, and (3) a monitor to assess task success and regulatory compliance. Notably, we introduce a VLA interface that accepts image-based visual targets from the VLM, enabling precise, goal-conditioned control. Our system successfully executes both primitive actions and complete multi-step chemistry protocols. Results show significant improvements in both success rate and compliance rate over state-of-the-art VLM and VLA baselines, while also demonstrating strong generalization to objects and tasks. Code, data, and models will be released.

Learning bimanual manipulation is challenging due to its high dimensionality and tight coordination required between two arms. Eye-in-hand imitation learning, which uses wrist-mounted cameras, simplifies perception by focusing on task-relevant views. However, collecting diverse demonstrations remains costly, motivating the need for scalable data augmentation. While prior work has explored visual augmentation in single-arm settings, extending these approaches to bimanual manipulation requires generating viewpoint-consistent observations across both arms and producing corresponding action labels that are both valid and feasible. In this work, we propose Diffusion for COordinated Dual-arm Data Augmentation (D-CODA), a method for offline data augmentation tailored to eye-in-hand bimanual imitation learning that trains a diffusion model to synthesize novel, viewpoint-consistent wrist-camera images for both arms while simultaneously generating joint-space action labels. It employs constrained optimization to ensure that augmented states involving gripper-to-object contacts adhere to constraints suitable for bimanual coordination. We evaluate D-CODA on 5 simulated and 3 real-world tasks. Our results across 2250 simulation trials and 180 real-world trials demonstrate that it outperforms baselines and ablations, showing its potential for scalable data augmentation in eye-in-hand bimanual manipulation. Our anonymous website is at: https://dcodaaug.github.io/D-CODA/.

Estimating the 6D pose of novel objects is a fundamental yet challenging problem in robotics, often relying on access to object CAD models. However, acquiring such models can be costly and impractical. Recent approaches aim to bypass this requirement by leveraging strong priors from foundation models to reconstruct objects from single or multi-view images, but typically require additional training or produce hallucinated geometry. To this end, we propose $\textit{UnPose}$, a novel framework for zero-shot, model-free 6D object pose estimation and reconstruction that exploits 3D priors and uncertainty estimates from a pre-trained diffusion model. Specifically, starting from a single-view RGB-D frame, $\textit{UnPose}$ uses a multi-view diffusion model to estimate an initial 3D model using 3D Gaussian Splatting (3DGS) representation, along with pixel-wise epistemic uncertainty estimates. As additional observations become available, we incrementally refine the 3DGS model by fusing new views guided by the diffusion model’s uncertainty, thereby, continuously improving the pose estimation accuracy and 3D reconstruction quality. To ensure global consistency, the diffusion prior-generated views and subsequent observations are further integrated in a pose graph and jointly optimized into a coherent 3DGS field. Extensive experiments demonstrate that $\textit{UnPose}$ significantly outperforms existing approaches in both 6D pose estimation accuracy and 3D reconstruction quality. We further showcase its practical applicability in real-world robotic manipulation tasks.

Active object localization remains a critical challenge for robots, requiring efficient exploration of partially observable environments. However, state-of-the-art robot policies either struggle to generalize beyond demonstration datasets (e.g., imitation learning methods) or fail to generate physically grounded actions (e.g., VLMs). To address these limitations, we introduce WoMAP (World Models for Active Perception): a recipe for training open-vocabulary object localization policies that: (i) uses a Gaussian Splatting-based real-to-sim-to-real pipeline for scalable data generation without the need for expert demonstrations, (ii) distills dense rewards signals from open-vocabulary object detectors, and (iii) leverages a latent world model for dynamics and rewards prediction to ground high-level action proposals at inference time. Rigorous simulation and hardware experiments demonstrate WoMAP's superior performance in a wide range of zero-shot object localization tasks, with a 63\% success rate compared to 10\%success rate compared to a VLM baseline, and only a 10 - 20\% drop in performance when directly transferring from sim to real.

Agile locomotion in complex 3D environments requires robust spatial awareness to safely avoid diverse obstacles such as aerial clutter, uneven terrain, and dynamic agents. Depth-based perception approaches often struggle with sensor noise, lighting variability, computational overhead from intermediate representations (e.g., elevation maps), and difficulties with non-planar obstacles, limiting performance in unstructured environments. In contrast, direct integration of LiDAR sensing into end-to-end learning for legged locomotion remains underexplored.We propose Omni-Perception, an end-to-end locomotion policy that achieves 3D spatial awareness and omnidirectional collision avoidance by directly processing raw LiDAR point clouds. At its core is PD-RiskNet (Proximal-Distal Risk-Aware Hierarchical Network), a novel perception module that interprets spatio-temporal LiDAR data for environmental risk assessment. To facilitate efficient policy learning, we develop a high-fidelity LiDAR simulation toolkit with realistic noise modeling and fast raycasting, compatible with platforms such as Isaac Gym, Genesis, and MuJoCo, enabling scalable training and effective sim-to-real transfer.Learning reactive control policies directly from raw LiDAR data enables the robot to navigate complex environments with static and dynamic obstacles more robustly than approaches relying on intermediate maps or limited sensing. We validate Omni-Perception through real-world experiments and extensive simulation, demonstrating strong omnidirectional avoidance capabilities and superior locomotion performance in highly dynamic environments.We will open-source our code and models.

Large-scale demonstration data has powered key breakthroughs in robot manipulation, but collecting that data remains costly and time-consuming. To this end, we present Constraint-Preserving Data Generation (CP-Gen), a method that uses a single expert trajectory to generate robot demonstrations containing novel object geometries and poses. These generated demonstrations are used to train closed-loop visuomotor policies that transfer zero-shot to the real world. Similar to prior data-generation work focused on pose variations, CP-Gen first decomposes expert demonstrations into free-space motions and robot skills. Unlike prior work, we achieve geometry-aware data generation by formulating robot skills as keypoint-trajectory constraints: keypoints on the robot or grasped object must track a reference trajectory defined relative to a task-relevant object. To generate a new demonstration, CP-Gen samples pose and geometry transforms for each task-relevant object, then applies these transforms to the object and its associated keypoints or keypoint trajectories. We optimize robot joint configurations so that the keypoints on the robot or grasped object track the transformed keypoint trajectory, and then motion plan a collision-free path to the first optimized joint configuration. Using demonstrations generated by CP-Gen, we train visuomotor policies that generalize across variations in object geometries and poses. Experiments on 16 simulation tasks and four real-world tasks, featuring multi-stage, non-prehensile and tight-tolerance manipulation, show that policies trained using our method achieve an average success rate of 77%, outperforming the best baseline which achieves an average success rate of 50\%.