Human Data Weakness: Can Quadrupeds Learn Backflips?

KyleMorgenstein · x · 2026-08-26

Discusses the limitations of human data in training generalist policies, specifically how a quadruped might infer complex maneuvers like a backflips without generating specific data from single-task experts. Argues that distillation from sim-trained single-task policies differs fundamentally from large-scale behavior cloning with human data, as the former doesn't scale as easily.

Related event: Debate: Sim Distillation vs Scalable Human Data in Robot Learning(2 posts)→

Original post →

More from Embodied

Embodied channel →