Quadrotor motion control using deep reinforcement learning

We present a deep neural-net-based controller trained by model-free reinforcement learning (RL) algorithm to achieve hover stabilization for quadrotor unmanned aerial vehicle (UAV). With RL, two neural nets are trained. One net is used as stochastic controller, which gives the distribution of control inputs. The other maps UAV state scalar, estimates reward controller. A proximal policy optimization (PPO) method, an actor–critic gradient approach, train nets. Simulation results show that achieves comparable level performance manually tuned proportional-derivative (PD) despite not depending on any model information. paper considers different choices function and their influence performance.