The industry's current main architecture
Brain and cerebellum: why a robot needs two «brains»
However strong a large model is, it cannot directly control a 1 kHz joint. Understanding the brain–cerebellum split is the highest-leverage piece of knowledge for entering robotics today — it explains where the algorithm, architecture, and embedded-control roles each sit.
One metaphor
The brain is navigation software: it tells you where to drive, and being a few seconds late is fine. The cerebellum is your own hands and feet: 1 ms late and the car is in the ditch. Safety reflexes are «pull the hand back when it burns» — they never pass through the brain at all.
The split in one table
| Layer | Typical instance | Frequency | What it owns | Latency requirement |
|---|---|---|---|---|
| Brain | VLA / multimodal large models | 1–10Hz | Understand the scene, parse the instruction, decide «what to do» | seconds, tolerable |
| Middle layer | Skill library / trajectory planning / WBC | 10–100Hz | Turn intent into executable trajectories and torque targets | stable within tens of ms |
| Cerebellum | Real-time control (joint servo / whole-body control) | 500Hz–1kHz (current loop 10kHz+) | Track trajectories, keep balance, close the torque loop | jitter < 100µs, above that it falls |
| Reflex layer | Hardware protection in MCU / hard limits | microseconds | Overcurrent, overspeed, e-stop — no model involved | must be hardware-level |
The table mirrors the human brain: the prefrontal cortex is slow but smart, spinal reflexes are fast and dumb. Not a coincidence — it is a division of labor forced by the physics of latency.
Three integration patterns
Cloud brain + onboard cerebellum
The large model runs in the cloud; the robot fetches instructions over the network. Fast to deploy, but a network drop is paralysis and latency jitter is uncontrolled.
Fits: demos and remote orchestration
Dual-system on board (current mainstream)
Slow system (VLA) + fast system (real-time control) both on the robot's own GPU, joined by one intent vector. Examples: Figure Helix, GR00T N1.
Examples: Helix · GR00T
End-to-end single model (research stage)
One large model maps pixels straight to joint commands. The most elegant, but high-frequency control latency and safety proofs remain unsolved.
Status: many papers, few products
Whatever the school, the bottom layer has one constant
Safety reflexes never pass through a model: overcurrent protection, hard limits, e-stop loops live in MCU hardware. The model may err; the reflex may not — which is why embedded control engineers are scarcer, not less needed, in the «large model era».
How to learn into each layer
| Layer you want | Matching role | Core topics | Entry |
|---|---|---|---|
| Brain layer (models) | Embodied-AI algorithms | C6 C7 C8 + F1 F2 F4 (data) | C8 · model survey |
| Middle layer (glue) | Systems software architecture | D8 D5 D6 + C-domain literacy | D8 |
| Cerebellum layer (real-time) | Embedded / control electronics | D4 D7 D11 D12 + E2 E3 E8 | D4 · PID lab |
One judgment: the brain layer's gate is papers and compute, the middle layer's gate is experience, the cerebellum layer's gate is measurement data — of the three, cerebellum people are scarcest and hardest to replace with a large model.