💡 The paper addresses the gap between current embodied agents and the need for truly closed‑loop learning during physical robot execution. Existing harnesses for embodied agents are largely open‑loop: they execute a fixed set of skills during a rollout and only reflect on performance after the episode ends. This post‑hoc reflection cannot influence the robot’s actions in real time, yet physical interaction requires decisions that track rapidly changing robot‑environment states at the frequency of each action. The authors argue that large end‑to‑end policy models are too slow to provide the necessary per‑step governance, and that a new approach is needed to evolve corrective behavior while the robot is acting.

Zetta pronounced “Zetta” is introduced as a closed‑loop embodied harness that keeps the base policy frozen but continuously evolves “runtime critics” and recovery skills online. The system operates on three separated time scales: 1 an action‑frequency loop that evaluates each action with a lightweight critic and can intervene immediately, 2 a rollout‑level loop that proposes new critic or recovery code after a full episode, and 3 a validation‑gated loop that tests proposed updates on a held‑out set before committing them to the harness. By generating code‑based critics and recovery modules, Zetta can modify behavior without retraining the underlying policy network. The authors also provide Z‑Infra, an infrastructure that separates the agent’s decision logic from heterogeneous execution resources, allowing the fast critic code to run on low‑latency hardware while the heavy policy runs elsewhere.

In experiments on two challenging robot benchmarks—LIBERO‑Pro and RoboCasa—Zetta achieves state‑of‑the‑art success rates of 90.8 % and 93.6 % respectively, using the same rollout budget as prior methods. The approach yields an 11.1‑times speedup in inference time because the per‑step critic is far cheaper to compute than a full policy forward pass. Success continues to improve as the system accumulates self‑exploration experience, and the learned recovery skills transfer zero‑shot to new tasks. The authors also report clear “Aha moments” where the robot discovers concise recovery strategies that were not present in the original skill set. Overall, the work demonstrates that a closed‑loop harness that self‑evolves code‑based critics and recovery skills can dramatically improve reliability and scalability of embodied physical intelligence.