Vision Based Autonomous Navigation of Quadcopter using Reinforcement Learning

This paper presents a vision based approach for autonomous navigation of quadcopter using deep reinforcement learning methods to learn control policies. This approach is based on model free methods and hence eliminates the need to explicitly model the agent and environment dynamics which can be a hard task due to hidden dynamics and complex structure. It has also been shown that the quadcopter learns to avoid obstacles while navigating through the environment to reach the destination point. A convolutional neural network which represents our learning policy has been trained on a variant of reinforcement learning method called deep Q-learning. The input to the network is the depth image of the front view of the quadcopter and the output control actions are commands to quadcopter on how to steer through the environment. Finally, this approach has been successfully tested in a virtual outdoor environment which shows the effectiveness of the proposed framework.