Robo-COP co-evolves robot orchestrators and VLA policies during deployment, lifting simulated held-out success from 64.8% to 73.8% and real-world success from 38.3% to 50.0%
Robo-COP co-evolves a robot system's VLM orchestrator and VLA policy during deployment: it curates skill-level demonstrations from its own executions (keeping successful skills even from failed episodes), fine-tunes the policy only when the data can address recurring failures, verifies each candidate on target skills, overall task success, and previously admitted skills before adoption, and revises memory built around the old policy; it raises mean held-out success from 64.8% to 73.8% across ten simulated RoboLab tasks and from 38.3% to 50.0% on three real-world tasks.