Merging of appearance-based place knowledge among multiple robots

If robots can merge the appearance-based place knowledge of other robots with their own, they can relate to these places even if they have not previously visited them. We have investigated this problem using robots with compatible visual sensing capabilities and with each robot having its individual long-term place memory. Here, each place refers to a spatial region as defined by a collection of appearances and in the place memory, the knowledge is organized in a tree hierarchy. In the proposed merging approach, the hierarchical organization plays a key role—as it corresponds to a nested sequence of hyperspheres in the appearance space. The merging proceeds by considering the extent of overlap of the respective nested hyperspheres—starting with the largest covering hypersphere. Thus, differing from related work, knowledge is merged in as large chunks as possible while the hierarchical structure is preserved accordingly. As such, the merging scales better as the extent of knowledge to be merged increases. This is demonstrated in an extensive set of multirobot experiments where robots share their knowledge and then use their merged knowledge when visiting these places.
QC 20200707