The authors describe a new framework based on Gaussian processes (GPs) to produce occupancy maps. The method combines the regression accomplished by classic GPs with the probabilistic least-squares classification algorithm, to associate observation data with distributions rather than well-settled points.

The authors are aware of some weaknesses in current occupancy map generators, among them the popular occupancy grid. Such limitations include the lack of interdependency among neighboring elements, the constraint on a predetermined scale in representing structures, and exaggerated memory requirements. Therefore, a neural network covariance function is used to evaluate dependencies between observations. The authors introduce an efficient way to substantially reduce the size of the training dataset, while still allowing accurate representations for larger environments at a lower computational cost. The training data is stored in a kd tree structure to accelerate the retrieval of neighboring objects.

Sections 1 and 2 provide an introduction and related work. The third section contains a very careful presentation of several facets of the method. The fourth section describes the experiments carried out during system evaluation on two types of simulated datasets. Another set of experiments operates on real-world data. Comparisons are made with the traditional occupancy grid in all the cases. The discussions in section 5 concern aspects such as application performance in the context of online use; the relation between the number of neighboring objects and computation complexity; and, as a future direction, handling dynamic objects. The last section (6) presents the conclusions.

The material is highly original, interesting, unambiguous, thorough, and beautifully illustrated with several graphical representations, such as variation plots to highlight the degree of uncertainty associated with the locations or outlines of unexplored regions. The paper is a valuable contribution to the field of robotics.