Tsinghua's Z-Trans Bets Embodied AI's Next Battle Is Post-Deployment Self-Evolution, Not Data Scale

新智元 · wechat · 2026-09-09

Embodied AI's focus is shifting from training stronger foundation models to making robots reliably complete tasks in the real world. Physical Intelligence CEO Karol Hausman recently argued that going from zero to 60% zero-shot capability is a data problem, but nobody has solved the last mile from 60% to deployable. Tsinghua AIR professor Cao Ting and IEEE Fellow Liu Yunxin have launched Z-Trans, open-sourcing a three-layer self-evolution stack: the Zeva model introduces In-Context Causal Learning, letting robots extract cause-effect relations from their own interaction traces without weight updates — lifting cumulative success rates from 26% to 73% on benchmarks, plus 10% more with human demos. The Zettaζ agent harness adds online closed-loop learning via action-level error correction, batched experience optimization, and validation-gated skill updates, hitting 90.8% on LIBERO-PRO and 93.6% on RoboCasa. Z-Infra provides the data, rollout and inference pipelines for full deployment-cycle evolution.

Original post →

More from Embodied

Embodied channel →