HT-Policy: multimodality isn't why diffusion policies win at imitation learning

chris_j_paxton · x · 2026-11-18

The common explanation for diffusion/flow-matching policies beating regression in robot imitation learning—better support for multimodality—may be wrong, argue researchers Yuchen and Maojiaoyuan. Real-world success actually hinges on stripping multimodality from robot data. They attribute the advantage to state-dependent variation and heavier-than-usual Gaussian tails, and propose HT-Policy, which learns an input-dependent residual scale to mitigate long-tail samples, making plain regression competitive with flow matching while being faster and simpler to train. Chris Paxton highlights the accompanying RoboPapers podcast on robot learning theory.

Original post →

More from Embodied

Embodied channel →