On June 19, 2026, Galaxy General Robotics officially released AstraBrain-WBC 0.5, the world's first humanoid general-purpose cerebellum GPT foundation model. This marks a historic milestone where the Scaling Law — long validated in large language models — is demonstrated for the first time in humanoid robot motion control.
The model was trained on approximately 2 billion frames (20,000 hours) of human motion data, covering dance, sports, daily activities, industrial operations, and collaborative tasks. Its action space coverage is 4-5 times broader than the widely used AMASS dataset. With 80.4 million parameters, it is the world's first humanoid real-time motion control model reaching GPT-1 scale.
AstraBrain-WBC 0.5 debuts a GPT-style causal Transformer architecture in robot motion control, redefining whole-body control as a continuous sequence prediction problem. Instead of outputting isolated joint commands, the model understands motion as a continuous 'language of movement,' similar to how GPT processes text sequences.
Key results from real-robot testing include:
- Zero-shot generalization success rate: 92.58% (vs 76.89% for traditional MLP architectures)
- End-to-end inference latency: just 0.39 milliseconds — 800x faster than a human blink
- 29-degree-of-freedom whole-body coordinated control
- Outperforms NVIDIA SONIC, TWIST, and Any2Track on multiple metrics
The paper has been accepted at CVPR 2026 and all code, models, and technical results are fully open-sourced. Galaxy General — with over 5.5 billion yuan in cumulative funding and a valuation exceeding 20 billion yuan — has positioned this as the 'cerebellum' component of its AstraBrain full-stack embodied AI architecture, which integrates brain (cognition), cerebellum (motion control), and neural control into a unified whole-body, whole-hand end-to-end large model.
