Robot Cerebellum

The cerebellum in the brain-cerebellum split: the layer that turns high-level commands into fast, stable joint motions.

Brain-cerebellum is an informal way China's embodied-AI industry describes a two-layer system, and this entry covers the cerebellum half. The brain is usually a multimodal large model that makes only a few decisions per second, responsible for understanding instructions and breaking down tasks; the cerebellum takes a sub-task like 'walk forward half a meter' or 'pick up the cup' and outputs joint commands tens to hundreds of times per second, while also holding balance and rejecting disturbances. It can be a reinforcement-learning-trained locomotion policy, a traditional controller such as MPC plus whole-body control, or a fast action module inside a skill library or a VLA. Companies define the boundary differently: narrowly, it covers only motion control like walking and balance; broadly, it also includes manipulation skills like grasping. BAAI's RoboOS calls its pluggable skill library the 'cerebellum skill library.'

ExampleFigure's Helix isn't labeled a 'cerebellum,' but it maps onto the same structure: System 2, which understands the scene and language, runs at 7–9 Hz, while System 1, which produces actions, runs as a separate real-time process outputting continuous commands for 35 degrees of freedom at 200 Hz.

Also called

Cerebellum, Motion Cerebellum, Cerebellum Skill Library

Related

Brain–Cerebellum Architecture · Dual-System Architecture (System 1 / System 2) · Hierarchical Control · Locomotion Control · Whole-Body Control · RoboBrain

Sources

RoboOS: A Hierarchical Embodied Framework for Cross-Embodiment and Multi-Agent Collaboration (arXiv:2505.03673)
Figure: Helix — A Vision-Language-Action Model for Generalist Humanoid Control

As of

2025-05