Robotics · Human–Robot Interaction
I'm a Ph.D. student of Shanghai Jiaotong University, advised by Prof. Cewu Lu and Prof. Chuan Wen. I got MEng degree from The Department of Mechanical Engineering at UC Berkeley, where I was fortunate to cooperate with PhDs and postdocs in MSC Lab led by Prof. Masayoshi Tomizuka.
Before my journey at Berkeley, I received my Bachelor's degree from School of Mechanical and Engineering, Qian Xuesen Class at Shanghai Jiaotong University. After graduation, I studied Finance at SAIF at Shanghai Jiaotong University and then dropped out.
I'm now building general physical intelligence at Noematrix Lab. I believe we cannot directly deploy fully autonomous robot from lab to home in the near future. Let's chat if you are also interested in human-robot interaction, agentic robotics, and policy steering.
Previously, I also interned at Prof. Bo Zhao's lab, Grasp Lab at Zhejiang University, advised by Prof. Huixu Dong, and robotics startup Maxinsights.
Human-Robot Interaction, Policy Steering, and Agentic Robotics. All about human.
arXiv, 2026
RoboPrompt lets users steer diffusion and flow-matching robot policies with sparse target points, drawn traces, and coarse directional guidance. A reusable module turns human intent into action drafts that the base policy refines, enabling intuitive failure recovery and policy improvement through steered rollouts.
Research Blog, 2026
RoboRSI combines multi-agent collaboration with Top-down Skill Refinement to develop reusable robot skills. It coordinates planning, execution, diagnosis, and local skill updates, enabling stable self-improvement for long-horizon mobile manipulation with human guidance and safety supervision.
Research Blog, 2026
Noe-0 co-designs data, models, and infrastructure for embodied foundation models. Non-embodied data in both pre-training and post-training preserves natural motion for dexterous manipulation, while joint action learning and implicit world modeling improve transfer across embodiments, objects, and scenes.
ActiveGlasses is a system for learning robot manipulation from ego-centric human demonstrations with active vision. To enable zero-transfer, we extract object trajectories from demonstrations and use an object-centric point-cloud policy to jointly predict manipulation and head movement.
CVPR 2026
Evo-1 builds on a native multimodal Vision-Language model (VLM), incorporating a novel cross-modulated diffusion transformer along with an optimized integration module, together forming an effective architecture. With only 0.77 billion parameters, Evo-1 achieves SOTA results on the Meta-World and RoboTwin suite.
RSS 2026 RoboData Workshop
U-Arm is a low-cost, universal, leader-follower teleoperation system for any commercial robot arms through three hardware configurations. Designed for researchers, educators, and robotics enthusiasts, it provides a standardized interface for diverse robot platforms.
Evo-0 introduces a plug-and-play module that implicitly injects 3D geometry features into VLA models by leveraging an off-the-shelf visual geometry foundation model.
Selected systems, prototypes, and things made for research or simply for fun.
A desktop robotic arm for smartphone photography hobbyists. Let the arm handle the camera moves while you focus on the shot. We developed the full stack in-house to bring the complete system's retail price below RMB 6,000. Tiny camera crew, tiny budget, surprisingly fluffy ears.
Elisabeth from Gintama, now hanging out on my desk. A little character-inspired desk ornament, made just for fun. No grand robotics mission here, just a very important desk companion.
A character in Apocalypse Hotel (末日后酒店). 3D-printed and controlled through a game controller.
Each finger has flexion and lateral DoF with torsion springs incorporated into every joint. The current configuration still has limited payload. Assembly remains a maddening job thanks to all the tendons.
An undergraduate class project. We designed turbines that adapt to both water and air, and integrated the system on top of an open-source flight control board. It hit a tree during takeoff, but we still believe it could fly.
