M5Stack Core2 runs tiny AI to balance robot 60秒

The project addresses a practical hurdle in physical AI: training directly on real hardware risks damaging the robot during trial and error. By first getting the pendulum to balance with a conventional PD controller, the author could safely collect 1,000 samples of stable behavior over about 5 seconds. Those samples then served as training data for a small neural network on a PC.

The network itself is deliberately minimal — two inputs (tilt angle and angular velocity), four hidden neurons, and one output (motor command). After training with scikit-learn's MLPRegressor, the weights and biases were transferred to the Core2, where a short MicroPython function performs inference. The author notes that this is closer to behavior cloning than to reinforcement learning; the AI reproduces the controller's input-output relationship rather than discovering a new balancing strategy.

Safety measures remained in place even during AI-only control: the PD command was computed in parallel, and the system stopped if the AI and PD commands diverged too much, if tilt exceeded 20 degrees, or if the control cycle became abnormal. The author also cautions that the disturbance magnitude was not measured, so the comparison between PD and AI disturbance rejection is not quantitative.