Creating and using probabilistic costmaps from vehicle experience

Probabilistic costmaps provide a means of maintaining a representation of the uncertainty in the robot's model of the environment; in contrast to the ubiquitous assumptive costmaps which abstract this uncertainty away. In this work we show for the first time how probabilistic costmaps can be learned in a self-supervised manner by a robot navigating in an outdoor environment. Traversability estimates garnered from onboard sensing are used in conjunction with colour information from a-priori available overhead imagery to extrapolate the traversability of locations previously traversed by the robot to a much larger area. Gaussian processes are used to predict the traversability at unknown locations in the 2D map, and a number of techniques to deal with heteroscedastic noise and varying confidence in the training data are evaluated. A prior technique to exploit the probabilistic nature of the map in a probabilistic heuristic for A* search demonstrates that planning over these maps can also be done efficiently.