Vision-based Drone Flight via Reinforcement Learning

This project was conducted within Robotics and Perception Group at Universitat Zurich.

Descriptions:

• Based on the observation of 2D bounding boxes from the YOLO detector and the drone dynamics from MoCap System, an RL-based drone controller was trained to achieve a stable and smooth following of the target drone.
• Designed a set of reward functions to achieve smoothness and robustness of the vision-based following. A PPO model was trained with a drone dynamic model, and experiments were conducted in both drone simulator and real world.