Horizon Open-Sources 400M-Parameter Robot Cerebellum Model HoloMotion-1, Enabling Various Complex Actions

Want robots to dance, work out, or even carry boxes? Horizon's newly open-sourced robot cerebellum model might make these possible.

Horizon recently released and open-sourced HoloMotion-1, a 400M-parameter model for full-body control of humanoid robots. Compared to traditional models of millions or tens of millions of parameters, it elevates the robot 'cerebellum' scale to a new level, achieving about 300FPS real-time inference on the edge.

Combining MoE sparse activation and KV-cache mechanism, the model reduces inference overhead while maintaining large capacity. Trained on multi-source motion data, it can zero-shot transfer to real robots, completing complex actions like large-range limb movements and dynamic balance.