Autonomous Exploration Under Uncertainty via Graph Convolutional Networks

We consider a mapping and exploration problem in which a range-sensing mobile robot is tasked with mapping the landmarks in an unknown environment efficiently in real-time. There are numerous state-of-the-art methods which consider the uncertainty of a robot's pose and/or the entropy and accuracy of its map when exploring an unknown environment. However, such methods typically use forward simulation to predict and select the best action based on the respective utility function. Therefore, the computation time of such methods is often costly, and may grow exponentially with the increasing dimension of the state space and action space, prohibiting real-time implementation. We propose a novel approach that uses a Graph Convolutional Network (GCN) to predict a robot's optimal action in belief space over a graph representation of candidate waypoints and landmarks. The learned exploration policy can provide an optimal or near-optimal exploratory action and maintain competitive coverage speed with improved computational efficiency.