Deep reinforcement learning based aggressive flight trajectory tracker

In this work, we aim to develop a deep reinforcement learning-based trajectory tracking controller for aerial vehicles based on the Proximal Policy Optimization (PPO) algorithm. The main goal is to minimize the error between the reference position and velocity from the trajectory and the aerial vehicle position and velocity at time t. We used an aerial vehicle dynamic model with PI for the attitude controller and PID for the attitude rate controller. A deep reinforcement learning agent is utilized for generating pitch and roll references to track a desired agile trajectory. Simulation results are conducted to compare our proposed solution with LQR and LQI based trajectory tracking controllers. Our approach has been applied to Crazyflie quadcopter to track the given agile trajectories with acceleration up to 6.2 m/s², while keeping the root-mean-square tracking error down to 5 cm.