Notmisha: imitation learning works when trained on the job — pretraining is what falls short

notmisha · x · 2026-09-18

Researcher notmisha argues that imitation learning does work if you keep training on the job — it's pretraining that falls short. The one-liner cuts into the current debate over robot/embodied AI training recipes: scaling pretraining data alone isn't enough, and continued learning in real tasks is where imitation learning pays off. A compact but discussion-worthy take.

Original post →

Notmisha: imitation learning works when trained on the job — pretraining is what falls short

More from Embodied