A research team has tested whether a general-purpose coding agent can control a robot using the robot’s existing software tools. Their system, OpenRUA, gives Claude Code—powered by Claude Opus 5—terminal access to ROS 2, a software system commonly used to connect robot programs with sensors and movement controls. The setup does not provide a custom workflow or special robot-control commands; it supplies ROS 2 documentation and basic tools instead.
The researchers report success rates of 99.0% on CaP-Bench and 87.0% on LIBERO-PRO. They describe this as zero-shot control: the agent was not given task-specific training. In OpenRUA, it treats perception as reading and processing files, and manipulation as writing and running code. The researchers do not provide benchmark settings or explain what counted as a successful task.
What did the agent do on its own?
The researchers also examined the programs the agent created during episodes. In 96.80% of episodes, it wrote programs to process raw sensor inputs and calculate measurements. In 95.87%, it built motion-control clients, such as code to operate a gripper. It wrote closed-loop control programs—ones that adjust movement using sensor feedback—in 50.13% of episodes.
Those observations matter because the approach leaves the agent to organize its work rather than directing it through a prescribed sequence. The researchers argue that an off-the-shelf coding agent can act as a visuomotor policy: a system that uses visual or other sensor information to guide movement. They do not say whether the benchmark runs involved physical robots, simulations, or both, and provide no hardware or task-count details.
What the paper does not establish
OpenRUA is presented as an alternative to earlier robot-use systems that, according to the authors, relied on complex custom harnesses, specialized workflows, or bespoke interfaces. Instead, it exposes the robot’s native ROS 2 interface and lets the agent write the code it needs. The authors say their code is available, though no usable link is provided here.
The results cover two named benchmarks and one coding-agent setup. They do not show how the system performs across different robots, in unscripted environments, or under safety constraints. Nor do the authors compare OpenRUA directly with other agents or control systems. The findings therefore show what this setup achieved in the reported tests, not that coding agents are ready to operate robots reliably in everyday settings.
Why it matters
For robotics researchers, OpenRUA offers a way to test coding agents through a robot’s existing ROS 2 interface instead of building a custom control layer first. The reported benchmark scores are promising for that setup, but the researchers do not report hardware, safety, or real-world performance.