Deep reinforcement learning based intelligent agent for smart autonomous system of drone in GPS-denied environments

Obstacle avoidance is a fundamental requirement for autonomous drones UAVs which operate in a cluttered and GPS-denied environment. This work explores a deep Reinforcement Learning (DRL) based approach for obstacle avoidance for UAVs in unknown environments. We don't assume a map of the environment a priori but rather rely on our sensor represented by laser scan is mounted on board the UAV. The robot controller achieves obstacle avoidance ability by pre-training in a simulation environment using ROS-Gazebo simulator. The system is based on the recent Deep Q-Network (DQN) framework where a convolution neural network structure was adopted in the Q-value estimation of the Q-learning method. Simulation results show the robustness to different kinds of environments. All of the experiments were performed using the same pre-training deep learning structure.