AI Overview
Solarian's AI stack sits on top of the standard Controller contract and gives a robot the ability to plan, decide, and act autonomously. It follows a dual-process design: a slow, deliberative planner decides what to do, while fast reactive behaviors handle how to do it.
The pieces
- Name
AI Controller- Description
A Controller subclass that runs a dual-process cognitive loop. A vision-language model (System 2) decides which behaviors to start or stop; the fast control loop (System 1) merges the actions those behaviors emit and forwards them to the robot. See AI Controller.
- Name
Skills- Description
Self-contained, restartable behaviors that run in their own process — a diffusion policy, a motion plan, a scripted routine. The AI Controller orchestrates them based on the planner's decisions. See Skills.
- Name
Policies- Description
The trainable models a Skill calls to produce actions (diffusion, ACT, VLAs, classical methods). Policies are trained and served independently of the Skill that consumes them. See Policies.
System 1 and System 2
- System 2 — the planner. Runs in its own process at a low rate (typically ~1 Hz). It reads the latest observation plus any user message, calls a VLM, and decides which Skills to start or stop.
- System 1 — the fast loop. Reactive Skills push the next
Actioninto a shared-memory queue. The controller drains the queue every tick and forwards a single merged action to the robot at control rate (30–500 Hz).
This split lets the robot be both smart enough to plan (slow, may fail, returns JSON) and fast enough to control (predictable, jitter-free).
Where to go next
- AI Controller — the orchestrator and its dual-process architecture.
- Policies — training and serving the models Skills call.
- Skills — the unit of behavior the AI Controller runs.