In the present work a neural net based implementation of the obstacle avoidance behavior of mobile autonomous robots is considered. Supervised learning from example-situations and the corresponding human reactions under those situations, are used to train single sided reflex actions of the robot. The use of perceptional symmetry allows perception-action mapping with a reduced sensor space dimensions. In an effort to mimic the actual process of obstacle avoidance behavior of human locomotion (or that of automobiles driven by human response) a combination of basic reflex actions and a higher level logical decision, is implemented. Simulation results are presented to show the effectiveness of the proposed strategy in typical obstacle situations.