# Robo‑COP lets robots improve themselves while deployed

> The new Robo‑COP system co‑trains robot policies and their orchestrators, boosting simulated success to 73.8% and real‑world success to 50%.

Oossa · 2026-10-08 · https://oossa.com/en/robo-cop-lets-robots-improve-themselves-while-deployed

Researchers introduced Robo‑COP, a method where a robot’s vision‑language‑action policy and its orchestrator are fine‑tuned together during real use. The system gathers its own mistake data, updates the policy, and only adopts the new version after it proves better on the tasks it was trained for. In ten simulated RoboLab tasks, success rose from 64.8% to 73.8%, and on three real‑world tasks it climbed from 38.3% to 50.0%.

## The facts

- Mean held‑out success in simulation increased from 64.8% to 73.8%
- Real‑world held‑out success increased from 38.3% to 50.0%

## Why it matters

It lets service robots get better on the job, cutting the need for frequent manual reprogramming.

## Sources & references

1. [Co-Evolving Robot Orchestrators and Policies through Deployment](https://arxiv.org/abs/2610.09228) – arXiv cs.RO, 2026-10-08

Last updated: 2026-10-08