We introduce a physically grounded interaction frame, an instantaneous local basis that separates force regulation from motion, and a method to recover it from demonstrations. Using this, Force Policy combines a global vision policy for free-space motion with a high-frequency local force policy that estimates the interaction frame and executes hybrid force-position control. In real-world contact-rich tasks, it outperforms vision-only and force-aware baselines, improving contact stability, force accuracy, and generalization to novel objects.

Propose an object-centric history representation based on point tracking, which abstracts past observations into a compact and structured form that retains only essential task-relevant information. Tracked points are encoded and aggregated at the object level, yielding a compact history representation that can be seamlessly integrated into various visuomotor policies. Our design provides full history-awareness with high computational efficiency, leading to improved overall task performance and decision accuracy. Our history-aware policies consistently outperforms both Markovian baselines and prior history-based approaches.

Propose DQ-RISE, which quantizes hand states to simplify hand motion prediction while preserving essential patterns, and applies a continuous relaxation that allows arm actions to diffuse jointly with these compact hand states. This design enables the policy to learn arm-hand coordination from data while preventing hand actions from overwhelming the action space. Experiments show that DQ-RISE achieves more balanced and efficient learning, paving the way toward structured and generalizable dexterous manipulation.

Develop AirExo-2, an updated low-cost exoskeleton system for large-scale in-the-wild demonstration collection. By transforming the collected in-the-wild demonstrations into pseudo-robot demonstrations, our system addresses key challenges in utilizing in-the-wild demonstrations for downstream imitation learning in the real world. Propose RISE-2, a generalizable imitation policy that integrates 2D and 3D perceptions, outperforming previous imitation learning policies in both in-domain and out-of-domain tasks, even with limited demonstrations. By leveraging in-the-wild demonstrations collected and transformed by the AirExo-2 system, without the need for additional robot demonstrations, RISE-2 achieves comparable or superior performance to policies trained with teleoperated data, highlighting the potential of AirExo-2 for scalable and generalizable imitation learning.

Propose FoAR, a force-aware reactive policy that combines high-frequency force/torque sensing with visual inputs to enhance the performance in contact-rich manipulation. Built upon the RISE policy, FoAR incorporates a multimodal feature fusion mechanism guided by a future contact predictor, enabling dynamic adjustment of force/torque data usage between non-contact and contact phases. Its reactive control strategy also allows FoAR to accomplish contact-rich tasks accurately through simple position control.

Introduce CAGE, a data-efficient generalizable robotic manipulation policy. With less than 50 demonstrations in the mono-distributed training environment, CAGE can effectively generalize to similar and unseen environments with different levels of distribution shifts (background, object and camera view changes), outperforming previous state-of-the-art policies. This work makes a step forward in developing data-efficient, scalable, and generalizable robotic manipulation policies.

Propose RISE, an end-to-end baseline for real-world robot imitation learning, which predicts continuous actions directly from single-view point clouds. Trained with 50 demonstrations for each real-world task, RISE surpasses currently representative 2D and 3D policies by a large margin, showcasing significant advantages in both accuracy and efficiency.

Develop AirExo, a low-cost, adaptable, and portable dual-arm exoskeleton, for joint-level teleoperation and demonstration collection. Further leverage AirExo for learning with cheap demonstrations in the wild to improve sample efficiency and robustness of the policy.

Collect a dataset comprising over 110k contact-rich robot manipulation sequences across diverse skills, contexts, robots, and camera viewpoints, all collected in the real world. Each sequence in the dataset includes visual, force, audio, and action information, along with a corresponding human demonstration video. Put significant efforts in calibrating all the sensors and ensures a high-quality dataset.

Propose a powerful AnyGrasp model for general grasping, including static scenes and dynamic scenes. AnyGrasp can generate accurate, full-DoF, dense and temporally-smooth grasp poses efficiently, and it works robustly against large depth sensing noise.

Introduce TransCG, a large-scale real-world dataset for transparent object depth completion, and a lightweight depth completion method DFNet based on the dataset.