AUTONOMOUS VISUAL NAVIGATION AND COLLISION-FREE STRATEGY USING DEEP REINFORCEMENT LEARNING

Tracked robots need to achieve safe autonomous steering in various changing environments. In this thesis, a novel end-to-end network architecture is proposed for tracked robots to learn collision-free autonomous navigation through deep reinforcement learning (RL). Specifically, this research improved the robot’s learning time and exploratory nature by normalizing the input data and injecting parametric noise in the network parameters. Three convolutional layers are used on the four
consecutive depth images for features extraction and then the features passed to the Dueling Double Deep Q-Network for calculating the Q-values.