Prof. Ting Cao’s group at AIR defines embodied self-evolution — a physical agent that keeps interacting in its deployment environment, corrects its own policy from feedback and accumulated experience, and steadily raises task success — and has built a full stack around it: Zeva for in-context causal learning, Zetta ζ for an online execution-and-recovery loop, and Z-Infra for end-to-end training and inference infrastructure. Zeva lifted a frozen model’s cumulative success rate from 26% to 73%; Zetta ζ reached 90.8% on LIBERO PRO and 93.6% on RoboCasa.
https://mp.weixin.qq.com/s/xhbrnIH3YEulimrzM3WVLA