Distributed inference-based multi-robot exploration

Abstract

This work proposes a technique for distributed multi-robot exploration that leverages novel methods of map inference. The inference technique uses observed map structure to infer unobserved map features. The team then coordinates to explore both the inferred and observed portions of the map. Individual robots select exploration poses by accounting for expected information gain and travel costs. Disputes are settled using local auctions of expected travel costs. The benefits of inference-informed exploration are demonstrated in both simulated explorations and hardware trials. The proposed technique is compared against frontier and information-based exploration approaches with varying numbers of agents and communication strengths. Map inference is evaluated using publicly available sensor datasets. The proposed inference technique improves the correctly estimated subset of the environment by an average of 34.47% (maximum 108.28%) with a mean accuracy of 95.1%. This leads to a 13.15% reduction in the cumulative exploration path length in the trials conducted.