Hugging Face has introduced Zetta, a novel closed-loop embodied harness designed to address the limitations of open-loop systems in physical robot execution. Unlike traditional harnesses that rely on fixed skills and post-episode reflection, Zetta governs actions in real-time, adapting to rapidly changing robot-environment states.
The system operates through three timescale-separated loops: action-frequency governance, rollout-level critic-recovery proposal, and validation-gated skill updates. This architecture allows Zetta to evolve code-based runtime critics and recovery skills online while keeping the base policy frozen, enabling continuous self-improvement without retraining the core model.
Zetta is complemented by Z-Infra, a rollout infrastructure that decouples agent logic from heterogeneous execution resources. Together, they achieve state-of-the-art success rates of 90.8% on LIBERO-Pro and 93.6% on RoboCasa, with an 11.1x inference speedup. The system's success scales with self-exploration experience, and learned skills transfer zero-shot to new tasks, with emergent robotic 'Aha Moments' observed.
These findings suggest that closed-loop harness self-evolution opens a promising scaling path for reliable physical intelligence, potentially transforming how robots learn and adapt in real-world environments.