air-embodied-brain/Zetta-Embodiment

Zetta is an efficient closed-loop embodied harness for self-evolving physical intelligence. It evolves code-based runtime critics and recovery skills online while keeping the base policy frozen. Project Page: https://air-embodied-brain.github.io/zetta/

What it solves

Zetta is designed to improve the success rates of embodied AI (robots) by allowing them to self-evolve their ability to recover from failures. Instead of retraining a massive base policy, Zetta evolves code-based "critics" and recovery skills online to handle specific failure modes, significantly increasing success rates on benchmarks like LIBERO-Pro and RoboCasa while maintaining high inference speeds.

How it works

Zetta uses a closed-loop "Evolution Protocol" to iteratively improve robot performance:

  1. Failure Clustering: It groups failed trajectories using video and telemetry to identify common failure patterns.
  2. Causal Diagnosis: A diagnostic stage explains the causal mechanism behind the failure.
  3. Candidate Generation: A "Candidate writer" creates a bundle consisting of a Critic (to detect the failure) and a Recovery actor (to fix it).
  4. Verification: New recovery skills are tested via "Shadow Replay" and a "Same-seed Gate" before being promoted to the final set of verified skills.
  5. Execution: During runtime, the base policy remains frozen, but the evolved Critic can propose a recovery action which a high-level decision authority (Role1) can accept or reject.

Who it’s for

This project is for researchers and developers working on embodied intelligence, robotic manipulation, and Vision-Language-Action (VLA) models who want to implement self-correcting behaviors without the computational cost of full model fine-tuning.

Highlights

  • High Performance: Achieves state-of-the-art success rates on LIBERO-Pro (90.8%) and RoboCasa (93.6%).
  • Efficiency: Provides an 11.1x inference speedup compared to traditional methods.
  • Broad Environment Support: Integrates with LIBERO, RoboCasa, and RoboTwin.
  • Modular Architecture: Separates the rollout runtime (inference/simulation) from the evolution logic (diagnosis/recovery generation).

Related

  • Project
    OpenBMB/DeepThinkVLA

    DeepThinkVLA is a Vision‑Language‑Action model that inserts an explicit chain‑of‑thought reasoning step before generating robot actions. Using a hybrid decoder (autoregressive reasoning → bidirectional parallel action generation), the authors train on a newly curated embodied CoT dataset and then refine with outcome‑driven reinforcement learning. The model achieves 97 % average success on the LIBERO benchmark, runs at only 0.175× the latency of a pure autoregressive baseline, and transfers zero‑shot to the newer LIBERO Plus suite.

  • Project
    gepa-ai/gepa

    GEPA is a framework for optimizing textual parameters like prompts and agent architectures using LLM-based reflection and Pareto-efficient evolutionary search.

  • Project
    Memento-Teams/Memento-Skills

    A skill-centric agent runtime that allows AI agents to design, execute, and persist reusable skills to improve their performance on complex tasks through a reflection-based learning loop.

  • Project
    RLinf/RPent

    RPent is an open framework for building embodied agents that use recursive interaction and memory to evolve their capabilities and improve success rates in long-horizon physical manipulation tasks.

  • Project
    InternRobotics/VLAC

    A Vision-Language-Action-Critic model that provides dense reward feedback and trajectory quality screening for real-world robotic reinforcement learning.