Spiced self-play accepted at CoRL: just 30 minutes of human data biases agents to right conventions

EugeneVinitsky · x · 2026-09-10

The spiced self-play paper has been accepted to CoRL. Eugene Vinitsky's takeaway: when skills can be learned from self-play, human data still matters — but only a little. Just 30 minutes of human demonstrations is enough to bias agents toward picking up the right conventions instead of converging on alien solutions.

Original post →

More from Embodied

Embodied channel →