Reproduce “Off-Policy RL with Online Inference Tuning: My Solution to the CoRL 2026 RoboEval Challenge” 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
Ilia Larchenko
The authors use off-policy reinforcement learning to refine a pretrained robot policy for RoboEval. Their loop learns from scored rollouts and tunes inference settings, while scripted examples support strategies absent from demonstrations or local exploration. In MuJoCo, the final policy reached RDS 0.990 on held-out seeds and 0.987 on official seeds; effects are not isolated, and real-robot performance is untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.