Towards Functional Relation Learning with a Physical Robot

The objective of this project is to investigate the possibility of a robot learning a model for functional spatial relations through interaction with its environment. As a proof of concept I examine the two relations “location control” and “support” between pairs of objects, relations which are integral components of the spatial preposition “on”. A simulated environment with a model of a robot arm is set up using the OpenRAVE framework. The robot manipulates objects in the scene to test for each relation, generating data used to train a logistic regression classifier with the sparse Bayesian modelling approach. The results suggest that the functional relation learning with a physical robot is feasible, despite limitations caused by both the physics engine and by the difficulties of grasping.