Black Box Optimization of PID controllers for Micro Aerial Vehicles

When operating flying robots we repeatedly face the necessity of tuning controllers for particular electronic and physical configurations. In the spirit of embodied robotic learning we want to tune the controllers during the system’s closed-loop operation. We present an experiment applying a black box Bayesian optimization technique to the problem of tuning a hierarchical vertical position PID controller of an off-the-shelf Micro Aerial Vehicle flight controller and interpret it in the context of Policy Search.