BeingBeyond, the embodied AI startup founded by Peking University tenured associate professor Lu Zongqing, has released Being-M0.7, a latent world-action model that lets a humanoid robot walk and manipulate objects at the same time. The model is pretrained on more than 10,000 hours of raw human egocentric video and motion data before it ever sees robot trajectories.
The team's premise is that real robot demonstrations for whole-body loco-manipulation are scarce and costly to collect, while human first-person video and motion capture are plentiful. The catch is that human data is messy: some clips have only video, some only motion, and human movements cannot be executed directly by a robot.
Being-M0.7 handles this in three stages. Pretraining maps video and motion into a shared space and jointly predicts future visual states and body motion. Mid-training adapts that prior to the robot's camera view and body. Post-training adds a separate action expert that reads the frozen prior's predictions and outputs whole-body commands. Instead of predicting pixels, the model works on frozen DINO image features, and both humans and robots share a compact 22-dimensional, head-rooted motion representation. Missing modalities are handled with learnable placeholder embeddings rather than discarded. Data sources include Ego4D, Nymeria, HumanML3D and other public sets, subsets of BeingBeyond's earlier work and purchased data.
On the SIMPLE simulation benchmark, which covers six humanoid tasks at three difficulty levels, Being-M0.7 succeeded in 128 of 180 trials, or 71.1%, the best among the baselines tested. The paper lists 64.4% for Ψ0, 52.8% for π0.5 and 33.9% for GR00T-N1.6, though Ψ0 did better on one task and π0.5 on two others. Removing human-data pretraining cut the score to 48.9%.
On a Unitree G1 humanoid with dexterous hands, the model succeeded in 13 of 15 trials across three walk-and-grasp tasks, the same 86.7% as the strongest baseline, GR00T-N1.6, and ahead of π0.5 at 7 of 15 and Ψ0 at 3 of 15. In other words, the real-robot result is a tie, not a lead.
The team notes clear limits. Tests covered only short-horizon tasks, and the tracking accuracy of the whole-body controller used to collect demonstrations caps how precise the real-robot data can be.
SEE ALSO: BeingBeyond Releases Being-H0.7 Embodied World Model Trained on 200,000 Hours of Human Video · Robotera's VPP2 World Action Model Tops RoboDojo, Scores 58.5% Zero-Shot on Real ALOHA Arms