In this project, an end-to-end quadrotor motion planning problem in dynamic environments is studied, based on reinforcement learning algorithms. The planner policy is trained with two policy gradient algorithms: advantage actor-critic (A2C) and proximal policy optimization (PPO).

The dynamics identification of quadrotor Crazyflie 2.1 and its aggressive flip maneuver implementation are studied in this project. The dynamics study lies in an estimation of the quadrotor inertia, computed through a data-based parameter optimization problem.The flip maneuver is realized with two different control strategies: 1. A feedforward controller based on a simplified 2D flip model 2. A feedback tracking controller using geometric control defined on the 3D rotation group SO(3).