Autonomous Exploration with Expectation-Maximization

We consider the problem of autonomous mobile robot exploration in an unknown environment for the purpose of building an accurate feature-based map efficiently. Most literature on this subject is focused on the combination of a variety of utility functions, such as curbing robot pose uncertainty and the entropy of occupancy grid maps. However, the effect of uncertain poses is typically not well incorporated to penalize poor localization, which ultimately leads to an inaccurate map. Instead, we explicitly model unknown landmarks as latent variables, and predict their expected uncertainty, incorporating this into a utility function that is used together with sampling-based motion planning to produce informative and lowuncertainty motion primitives. We propose an iterative expectation-maximization algorithm to perform the planning process driving a robot’s step-by-step exploration of an unknown environment. We analyze the performance in simulated experiments, showing that our algorithm maintains the same coverage speed in exploration as competing algorithms, but effectively improves the quality of the resulting map.