Neurofuzzy learning of mobile robot behaviours

Abstract

The era of mobile robotics for use in service and field applications is gaining momentum. The need for adaptability becomes self evident in allowing robots to evolve better behaviors to meet overall task criteria. We report the use of neuro-fuzzy learning for teaching mobile robot behaviors, selecting exemplar cases from a potential continuum of behaviors. Proximate active sensing was successfully achieved with infrared in contrast to the usual ultrasonics and viewed the front area of robot movement. The well-known ANFIS architecture has been modified compressing layers to a necessary minimum with weight normalization achieved by using a sigmoidal function. Trapezoidal basis functions (B splines of order 2) with a partition of 1 were used to speed up computation. Reference to previous reinforcement learning results was made in terms of speed of learning and quality of behavior. Even with the limited input information, appropriate learning invariably took place in a reliable manner.