Learning spatial relations between objects from 3D scenes

Abstract

In this work, we learn a limited number of abstractions which can then be used to form preconditions for motor actions. These abstractions take the form of spatial relations amongst objects. We consider three “classes” of spatial relation: The objects either are separated from, on-top of, or inside each other. We have tackled this same problem in previous work (Fichtl et al., 2013). Here we report on recent improved results using a novel application of histograms to visually recognise a spatial relation between objects in the environment. Using this histogram based approach we are able to report a very high rate of success when the system is asked to recognise a spatial relation.