Learning subsumptions for an autonomous robot

Abstract

Our main objective was to develop methods for training neural network controllers to guide a Nomad200 mobile robot to any requested location (goal) within its work area and to avoid objects in its path. Four experimental studies are reported. The first two studies focussed on training different neural network controllers to implement each of the two behaviours: obstacle avoidance was implemented by training controllers on stimulus-response pairs collected during operation of the Nomad under control of a simple built-in behaviour module; and find-goal controllers were trained using the Nomad's odometry system. The final two studies concerned coordination of the two behaviours when the Nomad was required to find a location in an environment containing obstacles. In one of these studies, two neural network controllers, one for each behaviour, operated in a subsumption architecture under the direction of the sensory input. In the other study, a single neural network controller was trained to produce the two behaviours. Comparisons favoured the single controller in that it automatically incorporated coordination between the behavioral layers and allowed them to interact in a natural way. (3 pages)