I am a Ph.D. student in Computer Science and Technology at Peking University, advised by Prof. Mengyuan Liu. My research interests include robot manipulation, robotics, and embodied AI.
I received my master's degree from the University of Chinese Academy of Sciences, where I was advised by Sheng Xu, Yongsheng Ou, and Wujing Cao.
Selected recent work · Full list on Google Scholar
arXiv preprint, 2026
CORE learns an outcome encoder from action-free videos and aggregates successful terminal states into visual goal prototypes. Conditioning robot policies on these prototypes transfers geometric and contact constraints across embodiments without action labels, improving manipulation over language-only task descriptions.
IEEE Robotics and Automation Letters, 2026
An energy-consistent contact model encodes friction-cone and object geometry through an anisotropic metric, yielding smooth contact dynamics without complementarity constraints. Combined with contact-implicit MPC, it selects robust contact regions and controls object motion efficiently for dexterous manipulation.
arXiv preprint, 2026
TERM introduces the Eval-Actions benchmark and AutoEval models to assess execution quality and source authenticity. Spatio-temporal visual aggregation, kinematic supervision, and reasoning-based optimization enable evaluation beyond binary success and distinguish policy rollouts from human teleoperation.
AAAI Conference on Artificial Intelligence, 2026
MP1 combines 3D point-cloud observations with MeanFlow to generate action trajectories in one network evaluation. Classifier-free guidance improves controllability, while a lightweight dispersive loss separates scene features, enabling fast and accurate manipulation without iterative denoising.
arXiv preprint, 2025
GPA-RAM integrates grasp priors learned from existing task demonstrations with a spatial policy built on attention and Mamba. This pairing improves initial grasp precision and models task context efficiently, making fine-grained manipulation more reliable without collecting extra grasp-pose annotations.
IEEE Transactions on Industrial Electronics, 2024
A learning-based controller combines position, velocity, and force measurements from human demonstrations to generate joint angular acceleration commands for assembly. Automatic data labeling, Lyapunov analysis, and Lipschitz constraints support stable contact behavior across connector-insertion tasks.
2023 · IEEE International Conference on Real-time Computing and Robotics (RCAR)