Body Schema Learning

Abstract

Kinematic models are widely used in robotics to describe the mechanism of a robot. For example, the kinematic model of a manipulation robot is typically specified by the position of its joints, and the size and orientation of its links (Craig, 1989; Sciavicco and Siciliano, 2000). Kinematic models are usually derived analytically by a robot engineer and thus rely heavily on prior knowledge about the geometry of the robot. When such a model is applied to a real robot, its parameters have to be carefully calibrated (Gatla et al., 2007) to ensure a high accuracy, for example, using expensive calibration systems at the robot manufacturer’s site. As robotic systems become more versatile and are increasingly delivered in completely reconfigurable ways, there is a growing demand for techniques to learn kinematic models automatically. Ideally, such techniques would neither require human intervention nor costly calibration equipment. This capability does not only facilitate the deployment and calibration of new robotic systems but also enables robots to autonomously adapt their models when the kinematics change, for example, as a result of hardware failures or material fatigue. Furthermore, the intelligent use of tools also requires the robot to include a tool dynamically in its kinematic model (Nabeshima et al., 2006).