SE3 Labs releases 500 bimanual teleop demos behind its real-robot benchmark

paigeinsf · x · 2026-11-12

SE3 Labs published the setup behind its real-robot eval—10 episodes per task, milestone scoring, π0.5 fine-tuned on 500 teleop demos—and released the dataset on Hugging Face (CC-BY-4.0): 6.3 hours of successful bimanual YAM manipulation across 10 tasks, 30fps with 6 cameras and full joint/gripper state. Fine-tuned Dream-1 and more API-accessible models are next, with all runs viewable. Caveats: π0.5's score reflects only those demos, and it's an older open-weight model.

Related event: Physical Intelligence pits zero-shot LLM Fable against fine-tuned π0.5 on real robots(5 posts)→

Original post →

More from Embodied

Embodied channel →