Gauging error of pose acquired by vision systems in bin picking applications

Picking a part from an unorganized pile of parts requires an accurate vision system integrated with a robotic arm. A proper metric for gauging pose error is therefore indispensable. Pose error is a combination of an error in the position vector and an error in the orientation matrix. Pose errors of a system under test (SUT) are calculated by comparing the poses obtained with the SUT with those obtained using a ground truth (GT) system whose measurements are registered to the SUT coordinate frame. Typically, the position error is calculated as the length of a vector connecting the SUT and the registered GT positions, and the orientation error is determined as the angle of relative rotation between the SUT and the registered GT orientation. However, many parts processed in industrial bin picking applications have axial symmetry and for such parts, the orientation cannot be determined uniquely. This causes the common metric for orientation error to be ambiguous and misleading. We show that a better and more reliable orientation metric can be calculated as the angle between the axes of symmetry for a part in the SUT and in the registered GT coordinate frame.