Brain–Cerebellum Architecture

Splitting a robot's software into a 'brain' that plans and reasons and a 'cerebellum' that executes movement.

This is informal shorthand used in China's embodied-AI industry (大小脑, literally “big brain, small brain”) for a layered robot system, borrowed from the division of labor in the human brain. The “brain” is usually a large multimodal model (one that handles both images and text), responsible for understanding instructions, perceiving the scene, breaking a task into steps, and other high-level decisions. The “cerebellum” turns each sub-task into concrete motion: it might be a library of skills, a VLA (vision-language-action) policy, or a motion controller built with reinforcement learning or classical control, running at a higher frequency and closer to the hardware. Splitting the system this way lets each layer be trained and swapped independently, and lets one brain coordinate several different robots. The Beijing Academy of Artificial Intelligence's RoboOS (2025) uses this architecture, with RoboBrain as the brain and a pluggable cerebellum skill library for execution. China's Ministry of Industry and Information Technology used a similar “brain, cerebellum, limbs” breakdown of key technologies in its 2023 Guiding Opinions on Innovation and Development of Humanoid Robots.

ExampleA humanoid robot hears, “hand me the cup on the table.” The brain model breaks this into walking to the table, picking up the cup, and handing it over; the cerebellum's walking controller and grasping policy then carry out each step.

Also called

Robot Brain and Cerebellum, Brain-Cerebellum Hierarchical Architecture

Related

Dual-System Architecture (System 1 / System 2) · Hierarchical Architecture · Embodied Foundation Model · Locomotion Control · Whole-Body Control · RoboOS

Sources

RoboOS: A Hierarchical Embodied Framework for Cross-Embodiment and Multi-Agent Collaboration
FlagOpen/RoboOS (GitHub)
工业和信息化部：《人形机器人创新发展指导意见》解读（2023-11） (Chinese)

As of

2025-06