A 3-level neuro-fuzzy autonomous robot navigation system

Mobile robot control is a quite active research area. In this context the term `control\´ has a broad meaning that includes many different controls such as lower level motor control, and "behaviour" control, where behaviour represents more complicated tasks, e.g. obstacle avoidance or goal seeking. In this paper, a 3-neural network module system is presented, for three different aspects of mobile robot control. The first is a reinforcement learning neuro-fuzzy controller that takes on the low level control of the mobile robot, trying to avoid obstacles and head to a target, utilizing ultrasonic sensor readings (local navigation). The second module is a topologically ordered Hopfield neural network, which performs global navigation, using a sensor-built environment map. The third module is an associative memory Hopfield network, with a non-local learning rule, where environment maps can be stored. When triggered, this module tries to match and complete the so far observed part of the environment with one of the known maps. The performance of this 3-level navigation system was tested in computer simulation runs and proved to be efficient and robust. Aspects of computer simulation and computational requirements are discussed.