3D Registration for Verification of Humanoid Justin’s Upper Body Kinematics

Kurzfassung

Humanoid robots such as DLR’s Justin were designed to interact with unknown environ- ments and cooperate with humans. To ensure safety and mobility they are built with light-weight structures and flexible mechanical components. This light-weight design en- ables the robot to achieve a compliant behavior of the arm kinematic chain, with the draw-back of having a low positioning accuracy at the TCP (Tool-Center-Point) end-pose of the hand. Torque sensors are used to approximate the deflections of the joints and compensate the TCP end-pose errors. However, this approximation is insufficient due to remaining unidentified errors and flexibilities in the joints and links. The identification of a maximum TCP end-pose error is essential for object manipulation and path planning. This thesis proposes a verification routine for the identification of the maximum bounds of the TCP end-pose errors by using the on-board stereo vision system. The error identi- fication is based on the pose estimation of 3D point clouds of Justin’s hand using state-of- the-art 3D registration techniques. A model of the hand is required for pose estimation. The hand’s CAD (Computer-Aided- Design) model does not accurately reflect the data observed by the stereo vision system, therefore we generate models of the hand by registering multiple 3D point clouds. Be- cause the hand is a self-occluding object a full model is not achievable. Partial models are generated by creating subsets of overlapping point clouds. We proposed a method for the rejection of non-overlapping point clouds based on a statistical analysis of the depth values. The partial models are created by applying an extended metaview registration method to the remaining subset of point clouds. The registration of a random point cloud of the hand to the model generates an estimate of the TCP end-pose. This estimate is compared to the measured TCP end-pose obtained by computing the forward kinematics of the arm using the joint angle measurements. The verification routine consists in repeating this procedure with N -random poses. The bounds of the TCP end-pose errors are identified as the maximum translational and rotational er- ror from the set of random poses. To evaluate the errors identified by using 3D registration, we implemented a marker-based pose estimation method with the ARTrack IR Optical tracking system. This method re- quired a robust calibration of tracking targets to Justin’s head and hands. The errors estimated by using this tracking system are considered as the ground truth. Experimental results showed that the estimated errors by using 3D registration are consistent with this ground truth. Therefore, our proposed verification routine is suitable for the bound iden- tification of Justin’s TCP end-pose errors. The identified bounds can be used to adjust the obstacle clearance in path planning techniques. Unexpected collisions with the envi- ronment can be prevented by setting this clearance to a value higher than the maximum bound of the TCP end-pose errors.