What you are building
A rotary inverted pendulum: a rod hanging off the end of a motor-driven arm, which the controller has to swing up and then hold vertical. It is the standard demonstration problem for feedback control, because vertical is an unstable equilibrium — the only way to stay there is to keep moving the thing underneath.
Total cost is around £20 in parts. The equivalent teaching rig from a lab-equipment vendor is roughly two hundred times that.
The parts, and what each one does
Section titled “The parts, and what each one does”| Part | Job |
|---|---|
| NEMA17 stepper + TMC2209 driver | Rotates the arm. The stepper is commanded in acceleration, not position. |
| AS5600 magnetic encoder | Measures the pendulum’s angle, 12-bit (4096 counts per turn), over I²C. |
| Arduino Nano (ATmega328) | Reads the encoder, runs the controller, drives the stepper. |
| Printed enclosure, arm, pendulum | The mechanics. The pendulum link has a slot for a 2p coin as tip mass. |
The Nano is the interesting constraint. It has 2 KB of RAM and no floating-point unit, and it has to close a control loop fast enough to catch a falling pendulum. That constraint is what shapes the entire training pipeline.
Two ways to run it
Section titled “Two ways to run it”Tethered. The Nano acts as a low-level server over USB serial; a laptop reads state, decides an action, and sends it back. This is how the policy is fine-tuned on the real rig, and how any computer-side controller would run.
Standalone. The controller runs entirely on the Nano, no laptop attached. This is the end state, and getting there is most of what the documentation below is about — a network small enough to fit has to be distilled from a larger one, not trained directly.
Which path do you want?
Section titled “Which path do you want?”- I want to build one. Start at the bill of materials and work forwards. The build section ends with a rig that responds to test sketches.
- I have a rig and want it balancing. Go to the pipeline. It is ordered, each step is idempotent, and it tells you what to check before moving on.
- I want to understand the design decisions. The How it works section is the reasoning: what a transition is, what gets randomised and why, and the transport-delay work that made on-device deployment possible at all.
- I just want to see it work. The demo on the front page runs the deployed network in your browser.
A note on honesty in these docs
Section titled “A note on honesty in these docs”Several numbers in this documentation are measured, and are reported with the date and conditions of the measurement. Where something was tried and failed — training a tiny network directly with RL, for instance, which failed on hardware twice — that is written down too, because the failure is the reason the pipeline has the shape it does.
The scoring metrics used throughout deliberately avoid the observation the policy sees: an “is it upright?” check computed from the policy’s own inputs can be satisfied by spinning the arm continuously, and early evaluations were fooled exactly that way.