Zetta $ζ$: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
Xin Ding, Liang Mi, Mingzhe Huang, Zixuan Wang, Chao Zhang, Zixu Hao, Fu Chen, Xiangyu Li, Yikai Zheng, Yaoyu Guo, Weijun Wang, Kun Li, Hao Wu, Yunxin Liu, Ting Cao
cs.RO
2026-08-17
Zetta freezes the VLA and evolves code critics plus recoveries. LIBERO-Pro Goal goes 34.5%→90.8%, RoboCasa 73.6%→93.6%, 11.1× faster inference than RPent.
End-to-end VLAs stay brittle off the demonstration distribution. Embodied agents were supposed to close that gap by orchestrating policies, code, and tools. Most harnesses still run open-loop: follow a fixed skill, then reflect after the episode. Contact physics changes on a millisecond budget. Post-hoc notes cannot query the state at the failure instant, and they cannot try a different action while the episode is still live.
Large agent models cannot decide at action rate. Governance has to live in code that can run online.
Zetta freezes the base policy and evolves a harness of runtime critics, recovery playbooks, and tools. An orchestrator only adjudicates evidence; its logic is also frozen. The policy is π0.5 on LIBERO-Pro and GR00T N1.5 on RoboCasa.
Three timescale-separated loops run together. At action rate, code critics scan the trajectory and propose a failure plus a mode; the orchestrator must accept the evidence before a recovery runs, and control returns to the VLA only after contact is stable. At rollout-batch scale, failures are clustered by earliest observable divergence and diagnosed top-down: evaluation, critic, state, planning, recovery, then parameters. Higher layers win if they already fix the fault. At iteration scale, only patches that raise success and hold across rollouts are packed into a versioned skill bundle.
Z-Infra splits agent logic from hardware: environment workers and model workers, with π0.5’s VLM and Action Expert in separate processes and activations moved over CUDA IPC. The paper calls this the first rollout stack built for self-evolving embodied agents.
Evaluation uses held-out seeds. RoboCasa: 50 development seeds and 50 unseen seeds per task. LIBERO-Pro: evolve on 50 development seeds, report only seeds 1–20.
On LIBERO-Pro Goal, averaging the T and S perturbations, success moves from 34.5% to 90.8% (Goal-T 31.0%→92.5%, Goal-S 38.0%→89.0%). Across all 40 Goal plus LIBERO-10 settings the macro average is 32.00%→71.13%. LIBERO-10 remains hard: 50.0%→63.0% (T) and 9.0%→40.0% (S), with several tasks still at 0%.
On 18 RoboCasa Atomic-Seen tasks, frozen GR00T goes from 73.56% to 93.56%. Skills learned on PnP-Stove (pregrasp, regrasp, stable place) transfer zero-shot to sink, cabinet, and toaster: 64%→84% macro, or 58%→82%, 62%→80%, 72%→90% on the three targets.
Inside a single failure cluster the curve can jump. Goal-T2 (bottle in bowl) stalls at 15% after a symptomatic patch, then hits 95% once grasp retention is the critic. Goal-S6 (cream cheese) jumps from 5% to 90%. Those checkpoints are internal versions of one diagnosis cycle, not extra VLA training.
Z-Infra on 8×A100 raises valid throughput from about 1.7 to 35.1 episodes/min. Versus RPent, inference latency drops by the stated 11.1×; at concurrency 16, throughput is 12.8×.
This is a deployment-time scaling path that never touches weights: freeze the VLA, stack verified code skills. Transfer follows physical failure modes rather than source-task trajectories. For a team that already has a decent base policy and mainly lacks execution reliability, that is cheaper than another demonstration corpus. It does not raise the policy ceiling. Several long-horizon LIBERO-10 tasks still do not start.
Everything is simulated; real robots are listed as future work. The 90.8% figure is the Goal-suite average, not the 71.13% over all 40 settings; the abstract blurs that. LIBERO held-out is 20 seeds. Critics in sim can read collisions and contact forces that a real stack must supply. Evolution needs large parallel rollout, so the paper’s budget is hard to copy on one box. The SOTA claim is explicitly “under our current rollout budget.”