In-hand object pose estimation from kinematic data and tactile sensing

Kurzfassung

Robotic hands are designed with anthropomorphic properties to allow their operation in human environments. Nonetheless, the execution of everyday manipulation tasks still proves challenging. A common problem in pick and place applications is the unintended movement of the manipulated object during the grasp acquisition. This can be the result of inaccuracies in the initial knowledge of object pose or the measured joint values of the hand. Although the acquired grasp may still be stable, not being aware of the correct object pose can cause the placing of the object to fail. Typically, in robotic manipulation, the object pose is acquired visually. However, during the execution of a grasp, occlusions of the object by the hand hamper visual localization. This thesis proposes the use of kinematic data and tactile sensing for the localization of the object. The presented in-hand localization system uses a particle filter to estimate the pose of object. Hypotheses of the object pose are created and evaluated with a collision-based measurement model. Furthermore, the movement of the object is predicted from control data, utilizing a grasp model. The quality of the implemented localization method is validated in experiments with the DLR Hand-Arm-System.