We didn't train a policy. We asked Claude-code to write a controller, and it wrote an MPC.
All the attention is on VLAs: models that look at a scene and output motor commands. Here's another way to use AI in robotics.
A triple inverted pendulum: three free rods, one cart, ±20 N. Four degrees of freedom, three of them unstable.
Built from scratch in half a day on a standard Motorcortex App. Claude Code wrote the Python controller, designed the MPC (Model-Predictive-Controller) and tuned it against a Physics Simulation model running in real time inside Motorcortex. It solved the swing-up offline by collocation, with the rail ends and force limit as hard constraints. It vibe-coded the phase logic and the dashboard too.
Every command is a full network round trip through the MOTORCORTEX parameter tree: 1.99 ms, 503 Hz, measured.
The API is the whole trick. An LLM can't tune what it can't observe. That's why we ship MCP servers across all Motorcortex tools.
So far it's simulation. We're building the hardware now, with the same code, real friction, real encoder noise.
Motorcortex by Vectioneer