Reproduce “RHO: Your Coding Agent is Secretly a Roboticist” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Karim Elmaaroufi, Justin Svegliato, Sarunas Kalade, Graham Schelle, Sanjit A. Seshia, Matei Zaharia
RHO searches robot-control policies by having a coding agent edit and test multi-file repositories before deployment. The selected repository runs without LLM code generation in the control loop, though perception and action services remain. On held-out LIBERO-PRO perturbations, RHO reaches 45.0% across 3,000 trials, but only 12.2% on object-swap, limiting the result to this benchmark.
Key insights
- Shifting the computational burden of LLM-based code generation from test-time to training-time through evolutionary search enables the deployment of complex, real-time robot policies that execute in a single turn.
- Optimizing entire multi-file policy repositories allows for more comprehensive refactoring and the discovery of robust, modular control logic compared to narrower search spaces.
- Tool-enabled coding agents, guided by iterative reflection on live environment feedback, can autonomously discover advanced neurosymbolic heuristics that enhance policy generalization and robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.