# OpenRUA lets coding agents control robots without custom harnesses

> Researchers show a coding agent can steer a robot directly through ROS 2, hitting 99% success on a benchmark, using only terminal access.

Oossa · 2026-10-05 · https://oossa.com/en/openrua-lets-coding-agents-control-robots-without-custom-harnesses

A team led by Zhaoyang Chu released OpenRUA, a minimal interface that lets a general‑purpose coding AI talk straight to a robot’s native ROS 2 software. The system gives the agent only a terminal, ROS 2 docs, and basic tools. The agent then writes its own perception and control programs on the fly. In tests, Claude Code (running Claude Opus 5) solved 99.0% of tasks on the CaP‑Bench suite and 87.0% on LIBERO‑PRO, all without any task‑specific training or extra software layers.

## How OpenRUA works

OpenRUA treats sensor data as ordinary files and motion commands as code that the agent can edit. The agent creates programs to read raw inputs, compute measurements, and generate motion commands. It even builds closed‑loop controllers that adjust actions based on sensor feedback in about half of the episodes.

## The facts

- OpenRUA was submitted to arXiv on 1 Oct 2026.
- Claude Code using Claude Opus 5 achieved 99.0% success on CaP‑Bench.
- The same setup reached 87.0% success on LIBERO‑PRO.
- The agent wrote perception programs in 96.80% of episodes.
- It built motion‑control clients in 95.87% of episodes and closed‑loop controllers in 50.13% of episodes.

## Why it matters

For robot developers, OpenRUA shows that a generic coding AI can replace custom software bridges, potentially cutting development time and cost. It also means that existing robot fleets could be programmed by AI without rewriting low‑level drivers, though reliability in real‑world settings beyond benchmarks remains to be proven.

## Sources & references

1. [OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies](https://arxiv.org/abs/2610.02459) – arXiv cs.RO, 2026-10-05

Last updated: 2026-10-05