Visual-motor coordination in unstructured environments: A self-organization approach

This paper proposes a new method for coordinating a robot system where the manipulator controller is directly connected to a visual camera system. In general, this coordination should be carried out in terms of a nonlinear transformation from the image space to the joint-angle space. In our method, the goal is achieved by decomposing the whole nonlinear transformation into local linear transformations. This process is accomplished automatically by a learning technique based on a Kohonen-style self-organization network. To put the robot system to good use even in an unstructured environment, some virtual forces are introduced into the original algorithm for improving its ability in dealing with further increased complexity underlying in the transformation.