Not All Diversity Helps: Noisy Robot Demos Waste Policy Capacity

chris_j_paxton · x · 2026-11-19

Robotics researcher Chris Paxton echoes Firat Candemir: with current imitation learning methods, "more diverse is better" doesn't hold. If every operator does the same task differently, the policy spends its capacity modeling that noise instead of the actual work.

Useful diversity comes from different workplaces and objects — not five slightly different grips on the same pick. Consistent demos beat a pile of slightly different takes.

Original post →

More from Embodied

Embodied channel →