Neural networks for behavioral conditioning of mobile robots

摘要

Biological systems provide excellent examples of self-organization in the form of visual object learning and behavioral adaptation. In light of this, a control system for mobile robots which is reminiscent of the modular distributed architecture of the brain is developed. The system utilizes neural networks for learning and performance at all stages, from visual object recognition to behavioral conditioning. The system includes networks for early visual perception, pattern learning and object recognition, object associations, emotional states, behavioral actions, motor control, and behavioral conditioning. These networks are interconnected by a variety of hard-wired as well as adaptive pathways. The system demonstrates self-organizing, teacher controlled, and reinforcement learning paradigms, and integrates these into a system in which external events interact with internal emotional states to produce emergent behaviors. The system, which has been implemented on a mobile robot MAVIN (Mobile Adaptive VIsual Navigator), takes as visual input various patterns of light (i.e., objects), learns and recognizes these objects invariant to size, location, orientation, angle of gaze (foreshortening effects), and aspect on the visual field. The system associates reflex motor behaviors with certain learned visual objects. A robot eye motion system that can detect and actively track moving objects is used on MAVIN. The system adapts to smooth motions of targets and executes predictive (compensatory) saccades, or suddenly saccades to a new target appearing in the visual field. This system also includes a vestibular-ocular reflex in which the neck moves to compensate for excessive eye motions. Using visual error feedback, MAVIN learns to compensate for motor errors associated with body movements. Control of these adaptive motor behaviors is achieved using collections of Adaptive Linear Neurons - ADALINEs. These sensory-motor reflex associations are finally used to demonstrate classical conditioning, in which a new visual stimulus is associated with the behavior triggered by either the on-set (conditioned excitor) or off-set (conditioned inhibitor) of some other stimulus. This type of associative learning is achieved using a REcurrent Associative Dipole - READ Circuit. Extinction of a conditioned excitor, and the non-extinction of a conditioned inhibitor are also demonstrated. These various conditioning paradigms are well known in animal behavior studies.