Self-Learning Exploration and Mapping for Mobile Robots via Deep Reinforcement Learning
Mapping and exploration of a priori unknown environments is a crucial capability for mobile robot autonomy. A state-of-the-art approach for mobile robots equipped with range sensors uses mutual information as the basis for a cost metric, and reasons about how much information gain is associated with each action a robot can take while constructing an occupancy map from its range measurements. However, the computational cost of such an optimization scales poorly as the number of potential robot actions increases. We propose a novel approach to utilize the local structure of the environment while predicting a robot's optimal sensing action using Deep Reinforcement Learning (DRL). The learned exploration policy can select an optimal or near-optimal exploratory sensing action with improved computational efficiency. Our computational results demonstrate that the proposed method provides both efficiency and accuracy in choosing informative sensing actions.
PDF Abstract