Multi-sensor gripper positioning in unstructured urban environments using neural networks

The future of robotics is not limited to factories and homes, and is extending to robot-assisted urban search and rescue. This paper proposes a new application of neural networks in this emerging field of research. Specifically, a neural network with feedforward architecture using the backpropagation learning algorithm is implemented in order to determine the positioning of a robotic gripper that is used in emergency rescue operations. Three training functions using the backpropagation learning algorithm are explored in order to improve the speed and the accuracy of the neural network. They are gradient descent, gradient descent with momentum and gradient descent with momentum and variable learning rate. The performance of each approach is evaluated through simulation. This work is part of an overall effort in developing a team of intelligent heterogeneous mobile rescue robots at the Industrial Automation Laboratory of the University of British Columbia. The goal is to have robots with various capabilities perform cooperative tasks that can provide assistance in extracting humans from life threatening situations. Such tasks may include using multiple robots to search for humans in distress, cooperatively grasping and manipulating objects to assist humans, and constructing simple devices with multiple robots in order to transport a human to safety.