Tsinghua's Zettaζ enables real-time robot error correction, boosting success rates to 90%

机器之心 · wechat · 2026-08-18

Tsinghua AIR and Z-Trans introduced Zettaζ, a closed-loop self-evolving embodied intelligence system. By employing high-frequency Critic mechanisms, it allows robots to detect anomalies and trigger recovery actions during execution rather than post-failure reflection. Success rates on LiberoPro and RoboCasa benchmarks jumped from 34.5% to 90.8% and 73.6% to 93.6%, respectively. Researchers observed 'Aha Moments' where success rates suddenly surged after plateaus. The accompanying Z-Infra boosted effective rollout throughput by 20.6x.

Original post →

More from Embodied

Embodied channel →