Experiments in self-calibration of an autonomous mobile manipulator

Current autonomous mobile manipulators are very complex machines built with dozens of motors and sensors connected through feedback loops. On top of this first layer of real-time controllers, subsequent levels of software modules interact in many ways to create increasingly sophisticated behaviours. One of the initial requirements for all these elements to work properly, is that the robot is properly calibrated. The size of each link and the relative positions among them have to be estimated. Also, many sensors have intrinsic parameters that have also to be estimated in order to relate the incoming data to a common reference system. In this paper, this crucial problem has been studied for a new humanoid robot named Loki. Loki has been built in the Robotics and Artificial Vision(RoboLab) laboratory for research in social and service robotics. Two different optimization approaches have been explored, theLevenverg-Marquardt gradient-descent procedure and the Markov chain Monte Carlo Simulated Annealing algorithm. Both methods are compared under this high dimension self-calibration problem and the results are analyzed and compared. Finally, several strategies to continue research in this area and to achieve fully autonomous calibration and re-calibration procedures are described.