Research · Oct 28, 2026, 00:00 UTC

Rollback paper proposes world-model counterfactual rollbacks for robot policy improvement

The method rewinds to critical decision points in a world model, branches into alternative actions, and uses LoRA-based correction from real transitions. Reported gains over the strongest baseline are 10.67%, 11.11% and 11.85% on three real-robot tasks.

• Topicresearch
• Topicrobotics