Structured lighting to enhance global image feature sensitivity in a neural network based robot-positioning task

This paper presents some promising results in visually positioning a 5-DOF robot arm using neural networks. The novelty of the method is in the technique used to extract global image descriptors, i.e., using a projection of a grid pattern on the surface of the target to create artificial features that enhance the sensitivity of global image descriptors to perturbations of the robot arm in the vicinity of the target object. Experiments results comparing the performance of this method to passive lighting are presented. It is found that this grid projection results in a better generalization of the network in learning the required mapping as compared to using passive lighting.