Visual Servoing for Floppy Robots Using LWPR

We have combined inverse kinematics learned by LWPR with visual servoing to correct for inaccuracies in a low cost robotic arm. By low cost we mean weak inaccurate servos and no available joint-feedback. We show that from the trained LWPR model the Jacobian can be estimated. The Jacobian maps wanted changes in position to corresponding changes in control signals. Estimating the Jacobian for the first iteration of visual servoing is straightforward and we propose an approximative updating scheme for the following iterations when the Jacobian can not be estimated exactly. This results in a sufficient accuracy to be used in a shape sorting puzzle.