Learning in Robot Vision Directed Reaching: A Comparison of Methods

摘要

Four neural network algorithms were examined for their ability to adaptively associate stereo camera coordinates with joint positions of a three degree of freedom manipulator arm in a 3D reaching task. Given reasonable numbers of training exemplars for an implementation in real hardware, all networks trained to significant errors. Two secondary error correction procedures were then tested. Both further reduced errors, but one method that depended on continuous visual and proprioceptive feedback to train a small set of associative weights that correlated joint and camera velocities was especially effective in eliminating errors. Stereo pan, tilt and vergence information was used to direct ballistic reaching, but relative depth information, was used for the visual feedback of end-effector velocity in the second error correction method. 1 Introduction The problem addressed in this study is the one of directing the end-effector of a robotic manipulator arm onto a visually located target....