Learning Understandable Concepts for Robot Navigation

Abstract

Service robots have to communicate with their human users in order to get commands, give reports, provide information, and get help in cases of failure. Service robots are not autonomous, but this does not mean that they can do without intelligence. They need intelligence in order to convert high-level human commands into their internal procedures and to adapt their execution to the actual environment. The concepts in commands and plans must be anchored in the “body” of the robot and at the same time be understandable to the human user. This means that robot and user must agree in a particular situation what a concept refers to, even though — because of their different sensory systems and action capabilities — the concept is defined completely different by robot and user. The concept definitions of the robot should include sensing and action so that the concepts become precise and specific when applied to a particular situation. We call such concepts operational concepts.

We apply machine learning in order to establish the link between high-level, human-oriented concepts and low-level perception and action of a mobile robot. Using the method of inductive logic programming, sets of rules for increasingly abstract concepts are learned. The rules are used for planning and plan execution in order to fulfill a user-given goal such as “move along the door”. The first-order logic representation allows time relations to be handled easily. Other advantages are given by unification and instantiation, which allow information that is unknown at the time of planning to be propagated from the environment through the levels of abstraction. The rules express relations — rather than particular value ranges — that must be valid for sensory patterns and for actions.