Abstract
This paper presents a method to coordinate an uncalibrated visuo-motor system in a 3D space. In order to handle spaces occluded by obstacles, a 3-camera system and two related self-organizing maps (SOMs) are employed. The self-organizing maps are directly connected to the camera system and are trained to perform position control. Based on the visibility of targets given in the workspace, an appropriate map is adopted. The maps determine the joint angles of the manipulator which make the end effector reach the targets precisely, and make the manipulator take obstacle-free poses. The proposed learning method ensures that the manipulator moves smoothly and consistently in whole workspace even though we use two maps to control it. In our visuo-motor system, neither any priori knowledge about the manipulator nor the camera parameters is required. In addition, the system is robust to change in its geometry. Simulation results are presented to demonstrate the effectiveness of the proposed method and the robustness of the system.
