Deep Learning-based Pose Estimation for Mobile Manipulator Tasks

Navigation errors typically lead to target pose errors in mobile manipulators. Although cameras can be attached to the end of the robot to calculate the pose required for a task through marker recognition and hand-eye calibration, pose errors may still occur because of camera distortion. To address this issue, a deep neural network was employed to compensate for the estimated marker pose error and reduce the transformation error between the endeffector and camera. The proposed deep learning-based pose estimation method minimized the 3D pose errors by 62% compared to the existing methods. Furthermore, the target objects were moved to the target jig in all 50 pick-and-place tasks, with a tolerance of less than 1 mm.