Rollback backtracks world models to improve robot policies
Original title: Rollback: Experience Backtracking in World Models for Policy Improvement
Absolute success gains were 10.67%, 11.11%, and 11.85% on three real-world tasks versus the strongest baseline.
Rollback treats action-conditioned world models as counterfactual engines: it rewinds to critical decisions, tries alternative actions, and turns outcome contrasts into action-level supervision without a separate value model. A training system couples distributed world-model search with asynchronous real-robot execution, and real transitions update dynamics through a lightweight LoRA pipeline the authors call test-time physics grounding. On block extraction, ball tossing, and tabletop curling, success rose over the strongest baseline, and on two tasks it passed average human success within 8 hours from weak teleoperated starts.
- Gains were 10.67% on block extraction, 11.11% on ball tossing, and 11.85% on tabletop curling.
- Policies started from teleoperated demonstrations far below human level.
- Ball tossing and tabletop curling exceeded average human success within 8 hours.
- Real transitions correct world-model dynamics through a lightweight LoRA pipeline.
From the original article
1 min read at huggingface.co
Action-conditioned world models offer a scalable alternative to costly real-world interaction for improving robot policies. However, existing world-model-based methods largely treat world models as forward-only surrogates of the physical environment, leaving untapped a capability no physical environment affords: rewinding to past decision points to explore counterfactual futures.
By Jiayi Luo, Yang Zhang, Liliang Chen · huggingface.co
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.