Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
Zetta keeps the robot policy frozen while updating runtime critics and recovery skills during execution.
The paper says current embodied harnesses mostly reflect after a rollout, too late for fast-changing physical interaction. Zetta instead uses three separated loops for action-level governance, rollout-level critic and recovery proposals, and validation-gated skill updates. The authors report 90.8% success on LIBERO-Pro and 93.6% on RoboCasa under their rollout budget, plus an 11.1x inference speedup. HF Daily Papers' note
The paper says current embodied harnesses mostly reflect after a rollout, too late for fast-changing physical interaction. Zetta instead uses three separated loops for action-level governance, rollout-level critic and recovery proposals, and validation-gated skill updates. The authors report 90.8% success on LIBERO-Pro and 93.6% on RoboCasa under their rollout budget, plus an 11.1x inference speedup. HF Daily Papers' note
score 5