Kinematic parameter error estimation from vision-based relative pose measurements

The precision with which a robot can reach a desired pose defined in the workspace depends on how well the nominal kinematic model implemented in the robot’s controller matches the physical robot. The accuracy of a robot can be improved by determining the deviations of the kinematic parameters from their nominal values and compensating for these deviations.This thesis focuses on the development of a method for determining the kinematic parameter errors of a robot without the need for costly measurement hardware and time-consuming initial calibration of the measurement setup. The virtual closed chainmodel is developed as a means of using relative pose measurements to estimate kinematic parameter errors. It relates small deviations in the kinematic parameters to a difference in the end-effector pose of the robot when it is moved from an initial measurement configuration into another. Such a model enables camera-based pose measurements of fiducial markers in the robot’s vicinity to be used for determining the kinematic parameter deviations without needing to determine the exact location of the markers.The developed method is evaluated using a series of datasets encompassing varying levels of parameter deviations and measurement noise. The error model is shown to have good convergence and high accuracy even for large parameter errors. The feasibility of using camera-based relative measurements is verified for datasets with a sufficient number of measurements and sufficient camera resolution. By utilizing a dataset containing 100 measurements extracted from a simulation environment, the average end-effector positionerror could be reduced to 0.51 mm after calibration. Using an extended dataset with 300 measurements further reduced the average error to 0.49 mm.