OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies

OpenRUA. Robot-Use Agents Are Zero-Shot Visuomotor Policies. By Zhaoyang Chu and six others, from University College London and Carnegie Mellon University. Published on the first of October. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

Coding agents are reaching into the physical world by writing and executing robot control programs. Engineers have spent decades building a mature software stack for reaching sensors and controlling motion, and one might expect agents to use it. Prior work instead mostly engineers complex custom harnesses, prescribing specialised workflows and supplying bespoke interfaces. The question posed here is whether that extra harness engineering is necessary. The answer offered is no: an off-the-shelf coding agent, handed the robot's native interface, can serve as a zero-shot visuomotor policy.

The test vehicle is OpenRUA, described as a zero-abstraction harness. It bypasses bespoke abstraction layers and gives an off-the-shelf coding agent only terminal access to the robot's native software interface, ROS 2. The design is minimalist: the workspace is the harness. It offers ROS 2 documentation and basic tools, and leaves the agent to organise its own work, with no agentic workflow orchestrated for it. Inside that workspace, perception becomes file input and output, and manipulation becomes coding. Those two recastings are what let a coding agent work on a robot at all. It already knows how to read files and write programs, so seeing becomes reading and acting becomes writing code that drives the robot's own software.

With Claude Code powered by Claude Opus 5, OpenRUA reaches a success rate of ninety-nine percent on CaP-Bench, and eighty-seven percent on LIBERO-PRO. Both are reached without bespoke primitives and without task-specific training. Zero-shot here means exactly that: nothing in the agent was trained for these tasks, and nothing in the harness was written for them either.

Further analysis turns up behaviours that emerge under this minimal design. For perception, the agent spontaneously writes programs that process raw sensory inputs and derive metric measurements, in ninety-six point eight percent of episodes. For manipulation, it spontaneously builds motion-control clients, gripper control for instance, in ninety-five point eight seven percent of episodes. Closed-loop control programs, which adjust motion based on sensor feedback, appear in fifty point one three percent. None of these were prescribed. The workspace only offered documentation and tools, and the agent chose to build them.

The claim is bounded by its setup. The headline results come from one agent configuration, Claude Code powered by Claude Opus 5, on two benchmarks. The closed-loop behaviour shows up in about half of episodes, not in all of them. What is shown is that additional harness engineering was not needed here, for these tasks; the abstract does not claim it can never help.

The contrast is between building an interface for the agent and letting the agent use the one engineers already built. Given the robot's own software, a terminal and its documentation, the agent wrote the perception and control code it needed by itself. The code for OpenRUA is public on GitHub.