Tag
This paper investigates whether per-phase metric selection improves demonstration curation for behavior cloning policies. The authors find that phase-gated curation never outperforms global or uniform metric application, and the dilution of defect signals across phases explains the failure.