GMM-based Learning from Demonstration of Dynamic Reaching Motions for Elastic Joint Robots

Kurzfassung

Reaching for objects is a fundamental aspect of human motion and daily functioning. In robotics, conventional approaches lag in the ability to perform these reaching-and-grasping motions smoothly and continuously. Furthermore, traditional approaches are based on executing pre-planning motions. Perturbations of the target or the robot during the motion require re-planning, which is usually time-consuming and thus limits the reactiveness of the motion. In recent years, Learning from Demonstration (LfD) has gathered research efforts to overcome some of these challenges. LfD learns to take optimal actions even in unstructured environments based on the current state of the system. However, the LfD community mainly focuses on developing approaches for the well-established rigid joint robots. Compared to rigid joint robots elastic joint robots offer increased safety and performance and are thus well-suited for the fast reach-and-grasp motion. Due to their mechanical structure, elastic joint robot controllers require higher-order reference signals which is why state-of-the-art approaches for rigid joint robots cannot be transferred directly to elastic joint robots. This thesis extends a Gaussian Mixture Model (GMM) based LfD approach, the Stable Estimator of Dynamical Sytems (SEDS), to learn the object reaching motions. The proposed method provides not only the typical velocity reference signal but additional derivatives up to the jerk level for the elastic joint robots. Furthermore, additional constraints in the optimization are developed to prevent discontinuities of the reference signals. The method is successfully tested on the DLR robot neoDavid, an anthropomorphic robot with variable stiffness actuators. It is shown that the method successfully learned the characteristics of the demonstration data for fast-reaching motions. Additionally, the experiments show that providing additional control signals prevents an overshoot for highly dynamic motions.