One 30-episode dataset, 5 robot policies: Pi 0.5, GR00T 1.7, Dream-1, ACT head-to-head

BharatJain · x · 2026-10-27

BharatJain collected 30 demonstration episodes on a Hugging Face LeRobot SO-101 arm using the LeLab GUI, training a pick-and-drop Jenga task with a POV webcam plus USB wrist camera (15s per episode, varied brick/box positions). He will train five policies on the same dataset — Physical Intelligence's Pi 0.5, NVIDIA GR00T 1.7, Dream-1, MomloACT2 and ACT — and share results, with environment/lighting variation planned for the next batch. A earlier 7-episode run already showed a fine-tuned ACT outperforming zero-shot Dream-1.

Original post →