Skip to content

Overview

CleanRUA (Clean Implementations of Robot-Use Agents) provides single-file implementations of robot-use agents for research, following CleanRL. Each file contains a whole agent: its prompt, its tools and the interaction loop. The environments are plain Gymnasium environments.

Successful episodes. Top, RoboDojo head camera: delegate stacking bowls, direct stacking blocks and direct lifting scissors. Bottom, DexJoCo ego camera: delegate clicking the mouse on its pad, delegate hammering the nail and monitor lifting the bucket with the boxed food inside

  • Single file. One algorithm per file, readable top to bottom.
  • Plain environments. Tools and models live in the agent, not in the environment.
  • Official verdicts. Success always comes from the environment's own check.

Start with the installation guide for RoboDojo or DexJoCo, then read the algorithms overview.

Citing CleanRUA

@misc{cleanrua2026yuan,
  title        = {CleanRUA: Clean Implementations of Robot-Use Agents},
  author       = {Mingqi Yuan, Xudong Wu, Jiayu Chen},
  year         = {2026},
  howpublished = {\url{https://github.com/Agentic-Intelligence-Lab/CleanRUA}},
}