Application of reinforcement learning to RC helicopter control

A reinforcement learning system composed of a radial basis function neural network trained by actor-critic algorithm is applied to control a small radio-controlled helicopter, which is difficult since the helicopter is very sensitive to small turbulence. As a first step, we construct a simple but enough rich simulator of the target helicopter and train the learning system with it. It acquires a sensitive controlling policy for simple task after a sufficient training. We apply the same system for the real world validation in the future. It is expected the reinforcement learning system can adapt to real one with less efforts after initial training in the computer simulation.