Policies trained from data instead of hand-written controllers.
LLM-written 3D value maps for zero-shot manipulation.
