Dynamic Obstacle Avoidance for Quad-rotors Using Deep Reinforcement Learning

Robotic agents are becoming more prevalent in many settings, and their use in unstructured environments poses numerous challenges. This work focuses on obstacle avoidance which is a key task for safe drone deployment in many contexts. Avoiding obstacles represents a great challenge for the state-of-the-art in perception, control and navigation and this topic has been widely covered in the robotics literature. The existing solutions however usually make strong assumptions that limit the generality of the environment or make use of expensive and not widespread sensors.The recent development in the field of Reinforcement Learning showed the potential of these techniques to solve problems previously considered impossible. The purpose of this thesis is to test to which extent an approach using Deep Reinforcement Learning can cope with the task of drone dynamic obstacle avoidance. The proposed approach uses auxiliary tasks together with the Reinforcement Learning process to enforce fundamental prior knowledge in the agent. The idea is to combine the appealing capabilities of Reinforcement Learning for unstructured problem solving with the potential of unsupervised learning techniques for obtaining synthetic information representation.This solution showed good results dealing with both static and dynamic obstacles in highly three-dimensional scenarios with a significantly higher sample efficiency with respect to other Reinforcement Learning based approaches