A robot learns reaction timing using neural nets combining with physical models

To suit actual situations and to adapt to new environments, robots should learn by experience to perform dynamic analysis of a sequence of pictures. For research on adaptive dynamic analyses, the reaction timing of the robots has been chosen in an important parameter in dynamic analyses. For this purpose, a robot system has been built. The robot returns a ball which rolls to it in a table. An algorithm provided by the vision system, called the ruler, is used to locate the ball in the pictures. Combined with the physical dynamic model of the robot and the ball, a feedforward neural net can improve the reaction timing. The neural net was trained via supervised learning. With correction by the neural net, the reaction timing of the robot has been improved