Third-year Ph.D. student in Computer Science, IIIS, Tsinghua University
I am a third-year Ph.D. student in Computer Science at Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University under the supervision of Prof. Huazhe Xu. I obtained my bachelor's degree from Shanghai Jiao Tong University (SJTU) and won the Best Bachelor Thesis Award.
My research mainly focuses on robotic manipulation, with an emphasis on building generalizable, dexterous, and robust manipulation capabilities. I am also interested in tactile sensing.
During my undergraduate study, I was fortunate to be mentored by Prof. Cewu Lu and Prof. Junchi Yan. I am also a research visiting student at MIT and MIT-IBM Watson AI Lab under the supervision of Prof. Josh Tenenbaum and Prof. Chuang Gan.
Selected recent updates
* indicates equal contribution, † indicates project lead. Representative papers are highlighted.
TouchBridge unlocks contact-rich dexterous manipulation from human demonstrations alone. By preserving contact through fingertip touch, it enables autonomous robot policies fine-tuned with zero teleoperated robot data and 7× faster demonstration collection, achieving performance close to policies trained on robot-collected data without object models or simulation.
UniDex is a robot foundation suite to construct large-scale robot data from egocentric human videos, and a vision-language-action policy for universal dexterous hand control.
This work introduces TactAR, a low-cost teleoperation system with real-time tactile feedback through AR, together with Reactive Diffusion Policy, a slow-fast visual-tactile imitation learning algorithm for contact-rich manipulation.
3D Diffusion Policy shows how simple 3D visual representations can be incorporated into diffusion policies for more generalizable visuomotor imitation learning.
DIFFTACTILE is a physics-based and fully differentiable tactile simulation system built to provide dense and physically accurate tactile feedback for robotic manipulation.
SLIM is a compact 0.5B-parameter policy that learns action-grounded predictive latents from robot trajectories, enabling robust, low-latency language-guided manipulation with a lightweight Mixture-of-Transformers backbone.
X-Distill transfers visual knowledge from a large DINOv2 model into a compact ResNet-18 encoder, enabling data-efficient diffusion policies that outperform larger vision models and policies using 3D observations across simulated and real-world manipulation tasks.
Maniwhere is a generalizable framework for visual reinforcement learning that enables robot policies to transfer across a mixture of visual disturbance types.
ThinShellLab is a fully differentiable simulation platform tailored for interactions with diverse thin-shell materials and varying mechanical properties.
Robo-ABC enables zero-shot generalization to manipulate out-of-category objects without manual annotation, extra training, part segmentation, or pre-coded knowledge.
ArrayBot is a distributed manipulation system made up of a 16x16 array of vertically sliding pillars with tactile sensing for simultaneous support, perception, and manipulation.
This paper proposes a robust flexible handover approach that helps robots grasp moving objects using dynamic motion trajectories with a high success rate.
This work studies contrastive learning from a collective point-set matching perspective and formulates it as a form of inverse optimal transport.
This paper proposes Relative Entropic Optimal Transport (RE-OT), a new matching-based view for unbalanced classification and prior-aware visual learning.
Selected honors
Service. Reviewer for CVPR, ECCV, ICRA, CoRL, IROS, ICLR, NeurIPS, RA-L, TASE.
