Embodied agents are increasingly used to close the gap left by end-to-end policy models. Yet the agentic path has not realized closed-loop learning in physical execution: existing harnesses remain largely open-loop, following fixed skills during rollout and reflecting only after an episode completes. Such post-hoc reflection cannot govern execution as it unfolds, because physical interaction requires decisions to track rapidly changing robot-environment states at a frequency beyond today’s large agentic models. We present Zetta ζ, a closed-loop embodied harness that evolves code-based runtime critics and recovery skills online while keeping the base policy frozen. Through three timescale-separated loops, Zetta ζ provides action-frequency governance, rollout-level critic-recovery proposal, and validation-gated skill updates. Together with Z-Infra, a rollout infrastructure decoupling agent logic from heterogeneous execution resources, Zetta ζ achieves state-of-the-art success on LIBERO-Pro and RoboCasa under our current rollout budget, reaching 90.8% and 93.6%, with a 11.1 × inference speedup; success continues to scale with self-exploration experience; learned skills transfer zero-shot, and clear robotic "Aha Moments" emerge. These results show that closed-loop harness self-evolution opens a scaling path for reliable physical intelligence.
No VLA fine-tuning. No gradient updates on policy weights. Only the code-based execution harness evolves.
Learned runtime critics execute at action frequency, monitor physical state, and trigger recovery skills the moment execution deviates from the nominal distribution.
Failed rollouts are clustered by earliest observable divergence, causally diagnosed, and converted into candidate critic and recovery updates in code space.
Only candidates that pass historical regression and strict held-out generalization are admitted into the versioned skill memory.
Success is not always a smooth curve. Zetta ζ often discovers the key physical bottleneck and unlocks a sharp, discontinuous jump.
Let's look at a case. Round-change improvements emerge on RoboCasa tasks such as turning on an electric kettle, sliding a dishwasher rack, closing a toaster-oven door, and placing a coffee pot. The CoffeeSetupMug task provides a representative example:
The policy correctly identifies the mug and coffee machine and begins transporting the mug. As the robot approaches the machine, however, the absence of continuous checks for grasp stability and collision clearance causes the arm to collide with the environment and lose the mug. The policy fails to detect this local physical anomaly in time, resulting in task failure.
Zetta ζ evolves a contact-aware manipulation skill from the failure. Its critic continuously monitors grasp stability, object drift, and collision risk. When triggered, the recovery transports the mug while maintaining safe clearance and relinquishes control only after confirming stable placement.
Here's another case from LIBERO-PRO. Consider the task Push the plate to the front of the stove:
The rollout fails at Failure 0: after grasping the plate, the policy does not maintain a stable grasp and loses the object before completing the transport.
Critic 1 (retained-grasp) determines whether the plate has been securely lifted, while Recovery 1 (retained-object transport) takes over the transport and resolves Failure 0. The rollout then encounters Failure 1: the grasp is lost during transport.
Building on Critic 1 and Recovery 1, Critic 2 (carry-retention) detects grasp instability during transport, and Recovery 2 (robust-carry retry) stabilizes the grasp and reduces transport risk. Together, they resolve Failure 1 and enable successful task completion.
As cumulative critic–recovery mechanisms are added, task success steadily approaches the measured capability ceiling of the frozen policy.
The skills evolved by Zetta ζ are not task-specific trajectories. They are defined by object-relative geometry, contact, grasp retention, and progress predicates — physical invariants that transfer across related tasks.
Pregrasp, regrasp, and stable-placement skills learned on PnP-Stove improve PnP-Sink, PnP-Cabinet, and PnP-Toaster from a 64% macro-average to 84%.
Target localization, collision-aware approach, and stable-contact skills learned on TurnOffStove improve faucet, cabinet, and microwave tasks from a 64% macro-average to 80%.
Self-exploration in the environment is the source of learning data, so faster rollouts produce faster evolution.
| Benchmark | Setting | Frozen VLA | Zetta ζ |
|---|---|---|---|
| LIBERO-Pro | Goal (T) | 31.0 | 92.5 |
| Goal (S) | 38.0 | 89.0 | |
| LIBERO-Pro | LIBERO-10 (T) | 50.0 | 63.0 |
| LIBERO-10 (S) | 9.0 | 40.0 | |
| RoboCasa | 18 Atomic-Seen | 73.56 | 93.56 |
A failed grasp becomes a targeted recovery, then control returns to the nominal trajectory.
@misc{ding2026zettazetaefficientclosedloop,
title={Zetta $\zeta$: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence},
author={Xin Ding and Liang Mi and Mingzhe Huang and Zixuan Wang and Chao Zhang and Zixu Hao and Fu Chen and Xiangyu Li and Yikai Zheng and Yaoyu Guo and Weijun Wang and Kun Li and Hao Wu and Yunxin Liu and Ting Cao},
year={2026},
eprint={2608.16590},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2608.16590},
}