Off-policy RL entry scores 0.990 RDS on CoRL 2026 RoboEval
Original title: Off-Policy RL with Online Inference Tuning: My Solution to the CoRL 2026 RoboEval Challenge
A π0.5 policy trained with AWR, RECAP, generated demos and failure recoveries also scores 0.987 on official seeds.
An author reports a two-part solution for the CoRL 2026 RoboEval challenge, which scores eight bimanual tasks on staged success, safety, efficiency and arm coordination against human demonstrations. Off-policy RL on a π0.5-based policy that is also its value function, using AWR and RECAP plus Thompson sampling of about 20 inference parameters, reached 0.85 RDS on held-out seeds in two days and was not converged. Generated full-reward demonstrations and recovery episodes from the policy’s own failures raised the submitted policy to 0.990 RDS on held-out seeds and 0.987 on official seeds.
Unconfirmed: Scores are the author’s report of a submitted policy, not an independent leaderboard confirmation.
- The challenge covers eight bimanual tasks scored against human demonstrations.
- About 20 inference parameters are tried by Thompson sampling on every rollout.
- The generic loop reached 0.85 RDS on held-out seeds in two days.
- Submitted scores are 0.990 RDS on held-out seeds and 0.987 on official seeds.
From the original article
1 min read at arxiv.orgThe CoRL 2026 RoboEval challenge has eight bimanual tasks, and it ranks a policy not only on whether it succeeds but on how: every stage of the task, done safely, efficiently and with both arms in coordination, scored against human demonstrations. My solution has two parts. The generic part, carried over from my LeHome challenge solution, is off-policy RL on a $π_{0.
By Ilia Larchenko · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.