Phase-Localized Curation Fails to Improve Robot Demonstration Filtering

In robot learning from demonstrations, it's hypothesized that filtering demonstrations based on metrics applied to specific temporal phases of a task could be more effective than using a single global metric. This approach would involve segmenting trajectories, scoring each phase with a locally informative metric, and then aggregating. However, recent research tested this "per-phase" hypothesis on three contact-rich robot pick-and-place tasks. The findings showed a negative result: phase-gated curation never emerged as the best strategy and often performed worse than applying the same metrics uniformly or using a strong single global metric. The failure mechanism was traced to signal dilution. When a defect signal is concentrated in one phase, aggregating scores across all phases dilutes this critical information with uninformative scores from defect-free phases, leading to the selection of a poorer demonstration subset. The study concludes that practitioners should prioritize identifying a single, defect-informative metric over complex phase-based decomposition.