Galaxy General Releases World's First Humanoid Robot General Cerebellum AstraBrain-WBC 0.5, Validating Motion Control Scaling Law

Galaxy General Robot has officially released AstraBrain-WBC 0.5, a cerebellum foundation model for full-body real-time motion control of humanoid robots under its AstraBrain technology system. Trained on approximately 2 billion frames (about 20,000 hours) of human motion data, the model has 80.4 million parameters, making it the world's first humanoid robot full-body real-time motion control large model reaching the GPT-1 scale.

Data Scale and Diversity

• The training dataset is the largest in the industry, covering open-source datasets such as AMASS, LAFAN1, Motion-X++, PHUMA, and MotionMillion, supplemented by over 1,000 hours of self-collected data.
• Data covers diverse scenarios including dance, sports, daily activities, industrial operations, and collaborative handling, with an action space coverage approximately 4 to 5 times that of AMASS.
• The team proposed the Harmonic Motion Embedding (HME) method, clustering data into about 300 action clusters to achieve diversity-aware balanced sampling, preventing common actions from overwhelming long-tail actions.

Architecture Innovation and Scaling Law Validation

• For the first time, a GPT-style causal Transformer architecture is adopted, redefining full-body control as a continuous sequence prediction problem. Combined with a motion prior library of 384 action experts, a unified control model is formed through distillation training.
• Experiments show that as data scales from 2 million frames to 2 billion frames, the model's zero-shot tracking error continuously decreases, with success rate improving from 83.26% to 92.58%, validating the Scaling Law in motion control.
• Under the same data volume, the Transformer architecture achieves a Mean Per Keypoint Position Error (MPKPE) of 43.25mm, outperforming TCN's 56.15mm by over 30%.

Core Capabilities and Performance

• Full-body coordinated control: Achieves complex actions such as hand-foot coordination and center-of-gravity shifting on a 29-DOF robot.
• High-dynamic motion: Zero-shot execution of high-dynamic actions not present in the training set, such as basketball, boxing, dancing, and rolling over to stand up.
• Millisecond-level real-time response: After optimization with TensorRT and C++, inference latency is as low as 0.39ms on a single RTX 4090, with a control loop at 50Hz, approximately 5 times faster than TWIST.
• Robustness: Maintains stability in scenarios involving fast motion, center-of-gravity changes, and complex contact transitions, with model success rate improving as data scale increases.

Open Source and Industry Impact

• The paper, code, and model are fully open-sourced (Paper: arxiv.org/abs/2606.03985, Code: github.com/GalaxyGeneralRobotics/Humanoid-GPT).
• As a motion control foundation model, it can generate high-quality VLA operation data, lowering the training threshold for full-body control models.
• Supports real-time full-body teleoperation and complex motion tracking, suitable for scenarios such as emergency rescue and hazardous environment handling.
• Developers can quickly generate creative motion content like dance and performance based on the model, enabling real-time generation and deployment.

Comparison and Evaluation

• In zero-shot tests on the Unitree G1 robot, AstraBrain-WBC 0.5 achieved joint position errors on four unseen dance motions lower than or comparable to current strongest open-source trackers such as GMT, TWIST, and Any2Track.
• The research team noted that this work is the first to validate a GPT-like Scaling Law in humanoid robot motion control, marking the transition from 'single-skill training' to the era of 'motion foundation models' in robot motion control.