Institute for Physical AI @ BMI · a demonstration
The reflex from the last lesson keeps it upright, but something has to decide where to push. Here that decision is a neural network — a real one, flying the real airframe on the board right now. Click the board to shove it, and watch it recover. Then guess how big the brain that does this actually is.
Provenance. The policy is a 6→24→24→3 multilayer perceptron, behaviour-cloned from the Layer-2 controller and run closed-loop in the exact rigid body: position and velocity in, desired acceleration out, with the trusted attitude reflex underneath. It is released open. The demonstration — a policy you can watch fly the real physics on-device, and count the cost of — is ours.