Institute for Physical AI @ BMI · a demonstration
The rule from the last lesson fixes how thrust becomes speed — but something has to decide how hard to beat the tail to swim at a chosen pace. Here that something is a neural network — a real one, setting the tail-beat that carries the robot to each target speed on the board. Watch it settle. Then guess how big the brain that does this actually is.
Provenance. The policy is a 1→10→10→1 multilayer perceptron — 141 weights — behaviour-cloned from the swim core's own inverse: target cruise speed in, tail-beat frequency out. It is the learned motor command that turns "swim this fast" into how hard to beat the tail, against the quadratic drag it cannot escape. Released open; the demonstration is ours.