Animal learning models as robot controllers

Abstract

Robots can do a range of wonderful things, but they can also appear really stupid. I would like my autonomous, sensor-rich, robot to be able to: complete its task whenever possible, despite distractions and disabilities; learn the best, most reliable cues for success of the various task components; have sensible default actions whenever the situation is unknown; cope with an unpredictably changing environment; and pay attention whenever I want to contact it. Dreamland? At the moment. Yet animals can do these things, and they are not inherently more capable than robots. So why not use an animal model as a robot controller? This paper describes work on the implementation and testing of a model of animal learning in a robotic context. The model is outlined and its interesting features described. An example under-specification problem is given. Experimentation summarised here included a trial naive implementation on an autonomous mobile robot and extensive classical conditioning simulations on computer. More details are given of a simulation experiment to produce behavioural chains and unlearn an unsuccessful chain. Current work involving new robot implementations is outlined. The appropriateness of using an implementation of this model as a robot controller is discussed.