Abstract:
Robot manipulation remains a challenging task. One of the key problems is uncertainty of the robot’s kinematic parameters such as its joint angles and the extrinsic calibration of its sensors with respect to its body. Even when the robot is properly calibrated, unknown dynamics and underactuated degrees of freedom introduce errors that can cause it to totally fail at common manipulation tasks. Can the robot correct for these problems automatically just using its sensors? We propose a method based on Bayesian graph SLAM that uses a monocular RGB camera mounted on the robot arm’s end effector to automatically calibrate and track it without the use of special markers or fiducials. We show that with our method, we are able to recover key parameters of the robot and their uncertainties in real time as the robot moves. We also provide analysis of how the robot’s structure and parameters of its sensor affect the SLAM system’s ability to track and calibrate it.