ETD

• dynamic generalization
• MAV
• quadrotor
• reinforcement learning

Tracking a reference trajectory with a small quadrocopter is a very challenging task. Nowadays the states of the art in this task use classical control strategies like model predictive control(MPC).
In this work, we develop a reinforcement learning based approach for tracking trajectories. The model is trained completely in simulation. Furthermore, it is able to generalize the performance to different simulators, in particular on a simulator very close to the real world like RotorS. The model is also able to generalize to drones with different mass and arm length, outperforming the MPC. We show then that it is possible to develop a general neural controller that relies on full state information.