BeingBeyond’s Being-M0.7 learns humanoid motion from video and beats prior methods on Unitree G1

jiqizhixin · x · 2026-07-24

BeingBeyond introduces Being-M0.7, a latent world-action model for humanoid robots.

It learns from egocentric videos and human motion data instead of expensive pixel-level prediction.
The model predicts future states in a compact latent space shared by humans and robots.
After lightweight post-training on limited robot demonstrations, it reportedly beats existing methods on real-world Unitree G1 tasks, including whole-body locomotion and dexterous manipulation in difficult interaction settings.

More from Embodied