Biologically-motivated learning in adaptive mobile robots

Over recent years, the focus of artificial intelligence (AI) research has shifted from top-down simulation of high-level cognitive functions based on (essentially-ungrounded) symbolic computations to consideration of the way that intelligent behaviour can emerge in bottom-up fashion in systems situated in the 'real' world. Inspirations from neurobiology and concepts of neural information processing have been very influential in this 'new' AI: most usually in the guise of 'parallel distributed processing' approaches operating at a high level of abstraction, or the more detailed and biologically-realistic 'computational neuroscience' paradigm. In this paper, we describe a biologically-motivated approach to learning in situated robots based on the computational neuroscience paradigm. The mechanisms by which such learning occurs are habituation, sensitization and classical conditioning of the neural responses involved in basic, pre-existing ('hand-wired') reflexes. Emergent light-seeking and collision-avoidance behaviors are observed in an adaptive mobile robot embodying these learning principles, as a result of interaction with its environment.