NVIDIA's ADEPT Pre-trains Dexterity, Cuts Robot Task Learning Cost by 3x

KyleMorgenstein · x · 2026-08-22

NVIDIA introduced ADEPT, a pre-training and post-training paradigm that treats dexterity as a prior rather than a per-task cost. The system learns reach-grasp-reorient-transport once in simulation using RL, then post-trains specialists. These policies achieve zero-shot deployment on real robot hands using pixels and touch, requiring only 3B steps per new task compared to 9B from scratch.

Related event: NVIDIA Unveils ADEPT: Dexterity Pretraining Speeds Up Robot Learning by 3x(3 posts)→

Original post →

More from Embodied

Embodied channel →