Gaussian Process Autonomous Mapping and Exploration for Range Sensing Mobile Robots

A framework for incremental autonomous mapping and exploration of unknown environments using the recently emerged Gaussian Process (GP) continuous occupancy mapping techniques is proposed in this article. The technique is able to exploit structural dependencies present in the environment as well as handle sparse sensor measurements. A strategy based on mutual information surfaces between the current map of the environment and predicted sensor readings using a one-step look ahead and macro-action concept is used to generate the control actions. We present results using the publicly available Intel Research Lab dataset with maps generated with occupancy grid-based nearest frontier, GP-based nearest frontier, and the proposed GP and mutual information-based exploration techniques. Maps are compared using the receiver operating characteristic curve and the area under the curve to demonstrate the accuracy of the proposed incremental mapping technique. Statistical significance test using two-sample t-test demonstrates the effectiveness of the proposed technique in terms of map entropy reduction rate and produced map quality ($p \leq 0.05$).