Inverse Reinforcement Learning Control for Trajectory Tracking for a Quadrotor UAV

Abstract

The main purpose of this thesis is to imitate the demonstrations of a quadrotor UAV flown by an expert pilot. First, we collect a data set of several demonstrations by an expert for a certain task which we want to learn. We extract a representative trajectory from the dataset. Hidden Markov model (HMM) and dynamic time warping (DTW) are used for obtaining the trajectory. We extract the sequence of state and input data. But a direct use of the input data can cause the danger in stability. For that reason, a controller is required. We design a reinforcement learning controller with reward function of linear quadratic form. To track the extracted trajectory well, an inverse reinforcement learning algorithm is suggested. Using particle swarm optimization (PSO), the reward function that minimizes the trajectory tracking error is learned. With the simulation and experiment applied to a quadrotor UAV, the successful imitation result is presented.