Speaker

Jiazhi Yang is a Ph.D. student at CUHK MMLab and a research intern at OpenDriveLab. His research focuses on developing autonomous, adaptive, and scalable self-improving embodied agents, through the lens of physical interactions and world models. He has publication records at premier venues including NeurIPS and CVPR, earning recognitions such as NeurIPS Spotlight and CVPR Oral presentations.

Abstract

Robot policies can transcend reactive behavior by learning to anticipate future outcomes and contextualize past events. In this talk, I will present two research contributions towards this goal. RISE leverages compositional world models to forecast action-conditioned scenarios and generate learning signals for robot policies. NativeMem integrates efficient memory systems into vision-language-action policies by compressing visual history into memory tokens, enabling policies to better contextualize past events.

Video

Coming soon. Stay tuned. :-)