Regrasp planning through throwing and catching

Multi-fingered hands are the most sophisticated and complex robotic grippers which can be planned to exhibit highly versatile grasping tasks. A relatively new regrasp planning method is by throwing and catching objects, which cannot be considered as a quasi-static problem due to its high dynamism influenced by the gravity, impact forces, air resistance, friction, etc., and thus has rarely been researched. In this paper, regrasp planning is performed through throwing and catching of various objects by a five-fingered anthropomorphic hand attached to a PUMA 560 manipulator. The planner uses an MLP neural network for learning from past throwing and catching experiences of objects with 9 different geometries. For the test set, five distinct objects with different geometries, densities, and center of mass were designed and tested. Experimental results of throwing and catching these objects showed that the minimum and maximum number of throws for successful catching was either 2 or 3, among which three objects needed just 2 tries (throws) thanks to the correct prediction of the implemented neural network.