A Mobile Robot Guidance System Based On Three Neural Network Modules

A combination of three separate neural network modules is employed to deal with the mobile robot control problem, giving very promising results. At the lower level, a neuro-fuzzy architecture, trained by reinforcement learning, steers the robot such that to avoid obstacles, exploiting ultrasonic sensor readings, and head to a target, given the heading error. ln the second and most important level, a topologically ordered Hopfield neural network performs global path finding in real time, using an environment map arranged on the fly. The third level involves an ordinary Hopfield neural network, used as an associative memory, which tries to match and complete the currently observed environment with one of the stored ones. Computer simulation valida tes the efficiency of the approach and shows its potential benefits.