#Berkeley (1)

WARP-RM: Warp-Augmented Relative Progress Reward for Data Curation

WARP learns dense signed relative progress from successful demos via time-warp augmentations, then WARP-BC filters and reweights action chunks for behavior cloning. On bimanual T-shirt folding, throughput rises up to ~18× vs vanilla BC as suboptimal demos increase; in 512 paired sim bottle scenes it reaches 290 bottles/hr.