Robots Have "Aha Moment"! Zetta ζ Enables Closed-Loop Online Learning for Physical Agents

Most existing embodied intelligence systems are 'static and open-loop', capable of executing fixed skills but struggling to perceive anomalies in real time and proactively repair during physical execution. Reflection by large models often occurs after an episode ends, making it impossible to test alternative actions online, credit assignment over complete trajectories is difficult, and retrospective analysis cannot capture the precise state at the moment of failure.

The real physical environment changes rapidly, and current physical agents can only review the entire trajectory after task failure, lacking the ability to judge, correct, and learn in real time based on state changes at the execution site. This leaves robots without online feedback when facing common issues like excessive grip force or pose deviation, preventing adjustment.

The core of Zetta ζ consists of three closed loops at different time scales. The first loop, the Critic-Governed Action Loop, continuously monitors the physical state at the frequency of action execution, triggering recovery actions immediately upon detecting anomalies, and uses the 'VLA Re-entry Contract' to ensure control is only returned after the fault is cleared.

The second loop, the Rollout-Batch Candidate Optimization Loop, clusters and diagnoses failed trajectories after batch tasks, identifies the earliest time point deviating from normal states, and traces the root cause along the hierarchy of 'evaluation → critic → state representation → planning control → recovery → parameters', automatically generating candidate updates.

The third loop, the Validation-Gated Skill Update Loop, first runs historical regression tests on candidate updates, then validates on held-out data, and only writes updates that genuinely improve success rates and generalize into the skill memory bank, avoiding overfitting.