Galaxy Universal Robotics, a Chinese humanoid-robotics company, has open-sourced AstraBrain-WBC 0.5, a foundation model it describes as a general-purpose cerebellum for humanoid robots. Where a robot's brain handles high-level planning, the cerebellum is the layer that actually moves the body, and this one coordinates whole-body, real-time motion across 29 degrees of freedom while keeping the machine balanced. The company calls it the first humanoid full-body real-time control model to work at this parameter scale, and the number is the surprise: 80.4 million parameters, small enough to run in under 1.5 milliseconds on a single RTX 4090.
The model was trained on what the company says is the largest human-motion dataset of its kind, roughly 2 billion frames covering about 20,000 hours of movement. The data spans dance, sports, everyday behavior, industrial operations, and two-person collaborative carrying, the idea being that a controller exposed to that range of human motion learns general principles of moving a body rather than a fixed list of routines.
The headline result is zero-shot generalization. The company shows the model performing complex actions that were not in its training data, including basketball moves, boxing, dance, somersaults, and coordinated carrying with a partner, without being retrained for any of them. End to end, from motion capture to robot, the pipeline runs in under 20 milliseconds. Generalizing to unseen motions is the hard part of humanoid control, where most systems are tuned skill by skill, so a single model improvising new full-body actions is the claim worth watching.
What makes it more than a demo is that the paper, the code, and the results are fully open-sourced. That is the opposite of the prevailing direction, where the most capable robot foundation models are proprietary and tied to a specific platform, and it means outside researchers can actually test whether the zero-shot claims hold on their own hardware. A small model that runs on one consumer GPU also lowers the barrier to putting capable control on a real robot rather than a server.
The caveats are the usual ones for a launch like this. The numbers and the world-first framing come from the company's own announcement, the demonstrations are curated, and how well controlled-setting results carry to messy real-world tasks is exactly what reproduction will decide. But the shape of the bet is the interesting part, and it runs against the moment: not a giant proprietary brain, but a small, efficient, open controller that anyone can download and try. If the zero-shot generalization holds up, doing more of robot motion with fewer parameters is a more useful direction than doing it with more.