Selected work

Representative Papers

Dexterous Contact-Rich Manipulation via the Contact Trust Region

May 04, 2025

This paper addresses the fundamental tension between local dynamical modeling and trustworthiness assessment in contact-intensive dexterous manipulation. We propose the Contact Trust Region (CTR) framework, which—uniquely—explicitly incorporates the physical constraint of unilateral contact into trust region design, thereby overcoming the physical inconsistency inherent in conventional ellipsoidal Taylor approximations. Within CTR, we formulate a computationally efficient global contact planning paradigm that synergistically integrates model predictive control with local path stitching. Our method constructs a roadmap for bimanual manipulation using the Allegro Hand on a standard CPU within 10 minutes, with online inference completing in seconds—substantially outperforming reinforcement learning baselines. Comprehensive validation in high-fidelity simulation and on physical hardware (KUKA iiwa bimanual platform with Allegro Hands) confirms that the approach achieves a rare balance of dexterity, real-time performance, and physical plausibility.

4 citationsRead paper

Clebsch-Gordan Transformer: Fast and Global Equivariant Attention

Sep 28, 2025

Existing equivariant Transformers suffer from limited expressivity and generalization due to reliance on low-order features and local attention mechanisms. This work introduces the first globally SO(3)-equivariant attention mechanism grounded in irreducible representations of SO(3). It employs sparse Clebsch–Gordan convolutions to efficiently couple features of arbitrary order, reducing computational complexity to O(N log N). Crucially, it is the first to systematically integrate Clebsch–Gordan coefficients into the Transformer architecture, while enforcing strict SO(3) × S_N equivariance via weight sharing and group-equivariant data augmentation. Evaluated on n-body simulation, QM9, ModelNet, and robotic grasping tasks, our method consistently outperforms existing equivariant Transformers—achieving simultaneous improvements in prediction accuracy, GPU memory consumption, and inference speed.

0 citationsRead paper

Recursive Deep Inverse Reinforcement Learning

Apr 17, 2025

Real-time inference of opponents’ goals and policies in non-cooperative multi-agent settings remains challenging, as existing deep inverse reinforcement learning (IRL) methods are predominantly offline, require large-scale datasets, and rely on first-order gradient optimization—rendering them unsuitable for time-critical adversarial interactions. Method: We propose the first recursive online IRL framework grounded in the maximum-entropy principle. It employs a second-order Newton method to tightly approximate the upper bound of the Guided Cost Learning objective and integrates ideas from the extended Kalman filter to enable rapid, robust parameter updates. Contribution/Results: Our method enables online cost function refinement from a single-step observation, drastically improving convergence speed and adaptability. Evaluated on standard and adversarial benchmark tasks, it achieves superior cost function recovery accuracy and policy generalization compared to state-of-the-art IRL algorithms.

0 citationsRead paper

Hierarchical Equivariant Policy via Frame Transf

Feb 09, 2025

Existing hierarchical policy learning suffers from coarse-grained interfaces between high- and low-level agents and neglects domain symmetries inherent in robotic manipulation tasks, leading to poor generalization and heavy reliance on demonstration data. This paper proposes HEP, a hierarchical policy framework for complex robotic manipulation. First, it introduces the “frame-passing interface,” where the high-level policy’s output is directly defined as a reference coordinate frame for low-level execution, enabling geometrically consistent cross-layer coordination. Second, it is the first work to theoretically model and guarantee full-system equivariance in hierarchical RL, explicitly embedding rotational and translational symmetries. Third, it integrates group-action modeling, equivariant neural networks, and geometric coordinate transformations. Evaluated in both simulation and real-robot settings, HEP achieves state-of-the-art performance, significantly improving long-horizon task reasoning accuracy and fine-grained manipulation stability.

0 citationsRead paper

Coarse-to-Fine 3D Keyframe Transporter

Feb 03, 2025

Current keyframe imitation learning (IL) methods neglect inherent spatial symmetries in robotic manipulation tasks, resulting in poor sample efficiency and limited generalization. This work identifies and formalizes the *dual equivariance* of keyframe action policies—specifically, equivariance under both workspace translations/rotations and gripper–object relative pose transformations. We propose a coarse-to-fine SE(3) action evaluation mechanism that decouples translation and rotation modeling while preserving joint optimization. Building upon Transporter Networks, we introduce the 3D Keyframe Transporter, which integrates cross-correlation-based feature matching, dual-equivariant feature encoding, and hierarchical SE(3) pose search. Evaluated across multiple simulated manipulation tasks, our method achieves an average performance gain of over 10%; on real-robot experiments, it attains an average improvement of 55%—significantly outperforming state-of-the-art keyframe IL baselines.

0 citationsRead paper

Recent publications

Latest Papers

Clebsch-Gordan Transformer: Fast and Global Equivariant Attention

Sep 28, 2025

Existing equivariant Transformers suffer from limited expressivity and generalization due to reliance on low-order features and local attention mechanisms. This work introduces the first globally SO(3)-equivariant attention mechanism grounded in irreducible representations of SO(3). It employs sparse Clebsch–Gordan convolutions to efficiently couple features of arbitrary order, reducing computational complexity to O(N log N). Crucially, it is the first to systematically integrate Clebsch–Gordan coefficients into the Transformer architecture, while enforcing strict SO(3) × S_N equivariance via weight sharing and group-equivariant data augmentation. Evaluated on n-body simulation, QM9, ModelNet, and robotic grasping tasks, our method consistently outperforms existing equivariant Transformers—achieving simultaneous improvements in prediction accuracy, GPU memory consumption, and inference speed.

0 citationsRead paper

Dexterous Contact-Rich Manipulation via the Contact Trust Region

May 04, 2025

This paper addresses the fundamental tension between local dynamical modeling and trustworthiness assessment in contact-intensive dexterous manipulation. We propose the Contact Trust Region (CTR) framework, which—uniquely—explicitly incorporates the physical constraint of unilateral contact into trust region design, thereby overcoming the physical inconsistency inherent in conventional ellipsoidal Taylor approximations. Within CTR, we formulate a computationally efficient global contact planning paradigm that synergistically integrates model predictive control with local path stitching. Our method constructs a roadmap for bimanual manipulation using the Allegro Hand on a standard CPU within 10 minutes, with online inference completing in seconds—substantially outperforming reinforcement learning baselines. Comprehensive validation in high-fidelity simulation and on physical hardware (KUKA iiwa bimanual platform with Allegro Hands) confirms that the approach achieves a rare balance of dexterity, real-time performance, and physical plausibility.

4 citationsRead paper

Recursive Deep Inverse Reinforcement Learning

Apr 17, 2025

Real-time inference of opponents’ goals and policies in non-cooperative multi-agent settings remains challenging, as existing deep inverse reinforcement learning (IRL) methods are predominantly offline, require large-scale datasets, and rely on first-order gradient optimization—rendering them unsuitable for time-critical adversarial interactions. Method: We propose the first recursive online IRL framework grounded in the maximum-entropy principle. It employs a second-order Newton method to tightly approximate the upper bound of the Guided Cost Learning objective and integrates ideas from the extended Kalman filter to enable rapid, robust parameter updates. Contribution/Results: Our method enables online cost function refinement from a single-step observation, drastically improving convergence speed and adaptability. Evaluated on standard and adversarial benchmark tasks, it achieves superior cost function recovery accuracy and policy generalization compared to state-of-the-art IRL algorithms.

0 citationsRead paper

Hierarchical Equivariant Policy via Frame Transf

Feb 09, 2025

Existing hierarchical policy learning suffers from coarse-grained interfaces between high- and low-level agents and neglects domain symmetries inherent in robotic manipulation tasks, leading to poor generalization and heavy reliance on demonstration data. This paper proposes HEP, a hierarchical policy framework for complex robotic manipulation. First, it introduces the “frame-passing interface,” where the high-level policy’s output is directly defined as a reference coordinate frame for low-level execution, enabling geometrically consistent cross-layer coordination. Second, it is the first work to theoretically model and guarantee full-system equivariance in hierarchical RL, explicitly embedding rotational and translational symmetries. Third, it integrates group-action modeling, equivariant neural networks, and geometric coordinate transformations. Evaluated in both simulation and real-robot settings, HEP achieves state-of-the-art performance, significantly improving long-horizon task reasoning accuracy and fine-grained manipulation stability.

0 citationsRead paper

Coarse-to-Fine 3D Keyframe Transporter

Feb 03, 2025

Current keyframe imitation learning (IL) methods neglect inherent spatial symmetries in robotic manipulation tasks, resulting in poor sample efficiency and limited generalization. This work identifies and formalizes the *dual equivariance* of keyframe action policies—specifically, equivariance under both workspace translations/rotations and gripper–object relative pose transformations. We propose a coarse-to-fine SE(3) action evaluation mechanism that decouples translation and rotation modeling while preserving joint optimization. Building upon Transporter Networks, we introduce the 3D Keyframe Transporter, which integrates cross-correlation-based feature matching, dual-equivariant feature encoding, and hierarchical SE(3) pose search. Evaluated across multiple simulated manipulation tasks, our method achieves an average performance gain of over 10%; on real-robot experiments, it attains an average improvement of 55%—significantly outperforming state-of-the-art keyframe IL baselines.

0 citationsRead paper