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.
More from Embodied
- Giving the Bristlebot a Brain: First Agentic Micro-Robot That Senses, Predicts, and Acts — maier_ak · 2026-08-18
- Stanford releases VISTA semantic navigation repo — rsasaki0109 · 2026-08-18
- Zoox robotaxi spotted in SF: odd shape but superior handling of street ambiguity — bookwormengr · 2026-08-18
- Unitree unveils 'Superman' humanoid robot that jumps 2m, runs 28mph — Polymarket · 2026-08-18
- Mixing RTX 3060 and 5060 Ti for local 30B model inference? — MkGod · 2026-08-18
- Bedrock Robotics builds infrastructure bots, hires top equipment operators — Sethwinterroth · 2026-08-18