CARF: Contrastive Attraction-Repulsion of Failure-Guided Flow Matching
Updated
Updated · arxiv.org · Sep 18
CARF: Contrastive Attraction-Repulsion of Failure-Guided Flow Matching
1 articles · Updated · arxiv.org · Sep 18
Summary
Researchers have introduced CARF, a new framework for robot learning that leverages both successful and failed demonstrations.
CARF uses a progress-based scorer to identify which parts of failed trajectories should be imitated or avoided, improving policy learning.
Tested in both simulated and real-world tasks, CARF outperformed existing baselines, potentially increasing data efficiency and robustness in robotic manipulation.