Robust nonlinear set-point control with reinforcement learning