Learning strategies for grasping using evaluation of grasps based on empiric experiments of an anthropomorphic hand

Kurzfassung

Service robotics for household applications is receiving an ever growing interest. Most tasks require robots to be able to grasp a wide variety of objects. Apart from simple grippers, the online planning of a grasp is a computationally expensive tasks, for this, most systems are relying on databases for online use. The grasps provided by databases are often planned on discretized models and simulations but the feedback for experimental knowledge is impossible, or at best, strongly limited. A brute force approach to experimentally evaluate all the grasps is deemed to fail since it would take too much time. Therefore, a pre-selection of the most promising grasps is needed. The main challenge, however, is that the relevant metrics for robot success in grasp execution are not known. As a result, the grasp planner might generate grasps which promise high analytical quality but are failing in execution. In this thesis, methods are investigated to improve the generated grasps under the constraint that only a small subset can be experimentally evaluated. The key idea is to use the outcome of the experiments to learn the relevant metrics and use those in simulation to improve the generation process. First, grasp planning approaches as well as grasp quality metrics are reviewed. Methods and parameters for grasp learning are identified and discussed. The most promising learning approach is implemented and experimentally validated on the DLR Hand-Arm-System. This method improves the quality of grasp databases and gives feedback about the relevance of the implemented quality metrics for the robot success.