Researchers gave coding agents direct robot control

A recent research paper introduced OpenRUA, a system that lets AI coding agents control physical robots without complex custom software. Instead of building specialized bridges for the AI, the developers gave the agent basic terminal access to ROS 2, the standard operating system for robotics.

By treating robot perception as simple file reading and manipulation as writing code, the researchers enabled the AI to work autonomously. The agent organizes its own workflow using nothing more than standard documentation and basic tools.

When tested with Claude Code powered by Claude Opus 5, the agent performed tasks with high precision. It achieved a 99.0% success rate on the CaP-Bench test and 87.0% on LIBERO-PRO. These results show that an off-the-shelf coding assistant can function as a robot brain without needing specific training for every task.

The system demonstrated impressive spontaneous abilities. In nearly all tested cases, the agent figured out how to write programs for raw sensor data and motion control. It even built closed-loop programs to adjust movements based on real-time feedback in over half of the attempts.

Imagine you could simply tell a robot to organize your desk, and it just starts writing the necessary code on the fly to interact with its surroundings. Does this mean we are moving toward a future where robots learn to handle any task just by reading manuals?

However, the current setup has limitations. The system was only tested on specific benchmarks and relies on the robot already running compatible software. It does not replace the need for physical hardware that works correctly, as the AI is limited to the capabilities exposed through the ROS 2 interface.