Reinforcement learning in a rule-based navigator for robotic manipulators

This paper reports on a navigation system for robotic manipulators. The control system combines a repelling influence related to the distance between manipulator and nearby obstacles with the attracting influence produced by the angular difference between actual and final manipulator configuration to generate actuating motor commands. The use of fuzzy logic for the implementation of these behaviors leads to a transparent system that can be tuned by hand or by a learning algorithm. The proposed learning algorithm, based on reinforcement-learning neural network techniques, can adapt the navigator to the idiosyncratic requirements of particular manipulators, as well as the environments they operate in. The navigation method, combining the transparency of fuzzy logic with the adaptability of neural networks, has successfully been applied to robot arms in different environments.