Manipulation strategy learning for carrying large objects based on mapping from object physical property to object manipulation action in virtual environment

In order to perform the carrying task of large and heavy objects, the robot needs to acquire the proper manipulation strategy depending on the object property. The manipulation strategy which depends on unknown physical parameters such as mass and friction property is difficult to plan beforehand, so acquiring strategy by trial and error is effective. In this paper, we propose the learning methods employing the dynamics-simulated virtual environment because handling large and heavy objects on a trial basis in real world can be dangerous. By adding operations to virtual objects with versatile physical parameters and analyzing the object motion from the aspect of contact state and 2D motion, we generate the database of manipulation strategy. The robot evaluates the safety and efficiency of operation and determines the feasibility of strategy based on the learned skill by trying strategy-proving operation.With the proposed system, the robot gets the ability to learn and execute the manipulation strategy which is given manually in previous research. We showed that the physical parameters of objects are estimated correctly and a real robot acquires the manipulation strategy for carrying objects automatically though the experiments.