AI

Co-Evolving Robot Orchestrators and Policies through Deployment

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 05:18:24 UTC.

Briefings

Robo-COP is a system where an orchestrator and policy co-evolve during deployment to improve robot performance. It curates skill demonstrations from its own executions and fine-tunes policies only after improvements are verified. Across ten simulated RoboLab tasks, it raised mean held-out success from 64.8% to 73.8%, while on three real-world tasks it improved held-out success from 38.3% to 50.0%. The system turns deployment into a self-improving flywheel where each execution improvement generates better data for the next learning round.

Key points

• Robo-COP curates skill demonstrations from its own executions before fine-tuning the policy.
• The system raises mean held-out success from 64.8% to 73.8% across ten simulated RoboLab tasks.

Context

Vision-language-action (VLA) policies trained on large datasets fail to generalize in real-world deployment. Agentic robot systems use a VLM orchestrator to learn when to call policies, but the frozen policy remains a bottleneck. Robo-COP proposes co-evolving the orchestrator and policy during deployment to overcome this limitation.

Limitations

Results are limited to ten simulated RoboLab tasks and three real-world tasks. Fine-tuning without verification reached only 65.8% success. The system's ability to overcome policy failures remains unproven beyond the simulated and limited real-world tasks.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/11/2026, 05:18:02 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihu6njxgt5fwof2dlzlnbtlrliniypbtex2fgnyfivpzp2re2bqlm