A Simulation Environment for the Manipulation of Naturally Variable Objects

Abstract

One area where the use of robots has been impractical up to the present time, is where the objects they handle are of an irregular shape. Robots are now very effective in manufacturing industries where their precision operations can be preprogrammed to produce machined parts of known dimensions to required tolerances. However, it is difficult to use robot arms to manipulate objects that are irregular and unpredictable. For example, in the food processing industry it is necessary to carry out operations such as shelling seafood, or filleting fish. The major problems are caused by inconsistencies in size, shape and texture. This work describes the possibility of using adaptive robot controllers to learn the correct operations by trial and error. The adaptive element is provided by a modified CM AC neural network, which implements a kind of reinforcement learning to gradually improve the robots actions. Rather than build a physical robot to carry out such a task, it was felt that a cheaper and more effective approach would be to create a realistic computer simulation environment in which to test out these ideas. This avoids spending a large amount of effort trying to maintain a real robot, which may eventually turn out to be inadequate to successfully execute the tasks required of it. By building an effective model, we may learn about the desired characteristics of such a robot and at the same time have a re-useable system with which we may tackle similar problems. We describe the system basics and our current progress towards these goals.