Vision-guided Collision Avoidance Through Deep Reinforcement Learning

Collision avoidance is a crucial task in vision-guided autonomous navigation. Traditional solutions tend to be computationally expensive and difficult to adapt to new environments. In this work, we propose a novel collision avoidance solution for autonomous drones. Formulated under a deep reinforcement learning framework, our model relies on a pair of margin reward functions to ensure the drones fly smoothly while greatly reducing the chance of collision. Additional reward functions are designed to attract the drones to fly towards their destinations, as well as to follow predefined routes. Experiments using indoor simulation environments demonstrate the effectiveness of our overall design and the individual components.