Zetta ζ, a new closed-loop embodied harness, reaches 90.8% on LIBERO-Pro and 93.6% on RoboCasa with an 11.1x inference speedup. The system evolves code-based runtime critics and recovery skills online while keeping the base policy frozen.
Key facts
- 90.8% on LIBERO-Pro benchmark
- 93.6% on RoboCasa benchmark
- 11.1x inference speedup reported
- Base policy kept frozen during evolution
- Critics and recovery skills evolved as code
Zetta ζ, introduced via a HuggingFace paper announcement, tackles a structural problem in embodied AI: how to improve a policy after deployment without catastrophic forgetting or expensive retraining According to @HuggingPapers. The answer is a closed-loop loop where the critic and recovery skills — not the policy — are the adaptive layer.
The frozen-policy design
Most self-improvement schemes fine-tune the base policy, which risks overwriting hard-won skills. Zetta ζ instead freezes the base policy and evolves code-based runtime critics that detect failure modes during execution. When a critic fires, a recovery skill — also generated as code — intervenes. This separation means the policy's learned distribution stays intact while the system's operational envelope expands.
The benchmark numbers are strong but not unprecedented: 90.8% on LIBERO-Pro and 93.6% on RoboCasa. The more striking claim is the 11.1x inference speedup, which suggests the code-based critics prune or bypass expensive model forward passes at runtime. The source does not disclose the base policy architecture, training compute, or whether the speedup is measured end-to-end or on a single component.
Why the critic-as-code approach matters
Critics trained as neural networks are opaque and hard to update incrementally. Evolving them as code makes each improvement a discrete, inspectable artifact that can be version-controlled and reverted. This is a meaningful departure from the gradient-descent orthodoxy that dominates the field — it treats runtime adaptation as a software-engineering problem rather than a weight-update problem.
The tradeoff is generality. Code-based critics are brittle outside the distribution they were evolved on, and the paper's abstract gives no indication of how the system handles novel object geometries or unseen task families. The 11.1x speedup may also be contingent on the critic being cheap relative to the base policy, which would not hold for smaller base models.
What to watch
Watch for the full arXiv paper and whether the authors release code or a benchmark harness. The key metric to track is whether the 11.1x speedup holds on larger base policies, and whether the code-based critics transfer across embodiments without re-evolution.
