Rotary Inverted Pendulum

The parts estimate belongs to a particular small rig

The BOM, dated May 25, 2026, estimates approximately £20 in parts and lists an ATmega328P Nano, short-body NEMA17 motor, AS5600 encoder and driver alternatives. Tools are listed separately. That estimate is not a verified current delivered budget and should not be treated as a like-for-like commercial laboratory-system comparison. [S005]

The printed pendulum incorporates a UK two-pence coin as its tip mass. The print guide explains that omitting it changes the mass and inertia assumed by the model. Source geometry, physical assembly and learned dynamics therefore need to remain consistent. The archive contains STL and URDF files; the external Onshape document was not independently exported in this review. [S005] [S007] [S012]

Reconcile the driver configuration before using the examples

The TMC2209 and DRV8825 options are not interchangeable pin-for-pin, and their current-setting conventions differ. The first-power-on guide also describes 16 microsteps for TMC2209 and 8 for DRV8825, whereas the selected standalone sketch explicitly sets MICROSTEPS = 32 and documents the corresponding driver pin settings. This discrepancy is preserved as a build check rather than silently presenting the pages as one consistent procedure. [S005] [S006] [S010]

The standalone sketch assigns the step signal to the Nano's Timer1-compatible pin and waits for magnet detection before engagement. Its settling pose becomes the encoder reference. These details make wiring and initialization part of the controller configuration, not incidental setup. No current setting, motion envelope or protective behavior was tested here. [S010]

A weight file name describes hardware, not network width

Committed policy headers are named by rig, driver and microstep ratio. The default sketch selects the rig1 TMC2209 32-microstep header; another preserved header for rig2 declares a 24–16–16–1 network. The final 32 in its filename is the microstep ratio, not a 32-unit hidden layer. The source includes the numerical weights, but that does not establish the complete original teacher checkpoints, replay buffers or measured training logs. [S009] [S010] [S011] [S012]

The distillation guide describes behavior cloning followed by DAgger under the intended deployment transport. It reports that the larger floating-point student missed the Nano's timing budget. The selected sketch sets 50 Hz control, a four-action smoothing window, four observation frames and a maximum velocity of 3.5 rad/s. These must agree with the policy's configuration; the weight header alone does not encode the operating point. The standalone procedure measures actual loop rate, because a nominal setting alone is not proof that inference meets it. Upstream balancing results remain reported results, not catalog reproductions. [S008] [S009]

The repository's MIT license accompanies the preserved source package. External CAD, toolchains and dependencies still need their own checks. Reproduction remains contingent on a matched physical rig, policy, measured timing and source dependencies; none was installed, trained or flashed during this review. [S003] [S005] [S009]