Embodied AI & Robot Learning
Hi there! I’m Jingyuan, currently an undergraduate student at Fudan University. I work on vision-language-action models, reward modeling and reinforcement learning for robotics, world models for manipulation, and continual, self-improving robot learning. My long-term vision is to develop general-purpose and trustworthy robots that can perform a wide range of everyday tasks, continually adapt in new environments, and operate reliably and safely in the real world.
Outside of research, I’m also a big fan of football and baseball!
Under review at ICLR 2027
ReAP is a policy-decoupled runtime monitoring framework that combines visual failure detection with action-conditioned future-risk prediction in a failure-relevant latent space.
arXiv preprint arXiv:2607.04426, 2026
ACE-Brain-0.5 develops a unified foundation model for physical agentic AI, bringing perception, reasoning, action, feedback, and self-improvement into one embodied system.
Predicting robot actions as future wrist-camera poses, using changes in camera position and orientation to represent robot motion for manipulation.