BeingBeyond, a startup in embodied artificial intelligence founded by Professor Lu Zongqin from Peking University, has introduced the Being-M0.7 model. This latent world and action model enables a humanoid robot to move and manipulate objects simultaneously.
The model was pre-trained on over 10,000 hours of raw human video recordings and motion data, even before it saw robot trajectories. The team believes that demonstrations of real robots for complex locomotion and manipulation tasks are rare and expensive to collect, whereas human video recordings and motion capture are widely available.
However, human data is imperfect: some clips contain only video, others only motion, and human movements cannot be directly reproduced by a robot. The Being-M0.7 model addresses this issue in three stages. In the pre-training stage, video and motion are transformed into a common space where future visual states and body movements are jointly predicted.
In the intermediate training stage, this prior information is adapted to the camera view and the robot's body. Finally, in the post-training stage, a separate action expert is added, which reads the predictions from the frozen prior information and issues commands for the entire body. Instead of predicting pixels, the model works with frozen DINO image features, and both humans and robots use a compact 22-dimensional head-attached motion representation. Missing modalities are processed using trainable placeholder embeddings instead of being discarded. Data sources include Ego4D, Nymeria, HumanML3D, and other publicly available datasets, as well as subsets of early BeingBeyond work and acquired data.
Testing Results
On the SIMPLE simulation benchmark, which covers six tasks for humanoids across three difficulty levels, Being-M0.7 achieved success in 128 out of 180 trials, or 71.1%, which is the best result among tested baseline models. The study indicates that $\Psi_0$ showed 64.4%, $\pi_{0.5}$ showed 52.8%, and GR00T-N1.6 showed 33.9%, although $\Psi_0$ surpassed competitors in one task, and $\pi_{0.5}$ in two others.
When pre-training on human data was removed, the performance dropped to 48.9%. When tested on the Unitree G1 humanoid with dexterous hands, the model succeeded in 13 out of 15 trials across three walking and grasping tasks, corresponding to the level of the strongest baseline model GR00T-N1.6 at 86.7%, and surpassing $\pi_{0.5}$ (7 out of 15) and $\Psi_0$ (3 out of 15). Thus, the result on the real robot was equal to, rather than exceeding, the baseline.
The team notes obvious limitations. Tests were conducted only for short-horizon tasks, and the tracking accuracy of the full-body controller used for collecting demonstrations limits the maximum accuracy of the real robot data.