Obstacle avoidance using neural networks

A neural network that limits the closest point of approach of an autonomous underwater vehicle (AUV) with respect to a navigation obstacle is described. Neural network inputs consist of beam outputs from a forward-looking sonar, and differences between current and desired values for AUV course and speed are inputs to normal navigation and control. The neural network outputs are AUV rudder angle and propulsion power: basic vehicle maneuvering characteristics are incorporated in the model. Obstacle avoidance is accomplished using a proximity detector for avoiding static obstacles and a rate detector for avoiding moving obstacles. The detections are made using 2D masked binary filters implemented as multilayer neural networks in the classification mode. Adaptive training is not used: instead. neuron weights are defined by the desired AUV response. The AUV simulation successfully avoided collision with all obstacles during test runs