BeingBeyond's Being-M0.7 teaches humanoids to walk and grasp using 10,000 hours of human video

BeingBeyond, an embodied AI startup founded by Peking University associate professor Lu Zongqing, released Being-M0.7, a latent world-action model pretrained on over 10,000 hours of human egocentric video and motion data. It scored 71.1% on the SIMPLE simulation benchmark (128 of 180 trials) and 86.7% (13 of 15) on a Unitree G1 humanoid.

• Pretrained on 10,000+ hours of human video and motion data before robot trajectories
• SIMPLE benchmark: 71.1% success vs 64.4% for Ψ0 and 52.8% for π0.5
• Removing human-data pretraining drops the score to 48.9%
• On Unitree G1: 13 of 15 trials, matching GR00T-N1.6