Learning to Grasp from Examples in Telerobotics

C. Fernández, M.A. Vicente, C. Pérez, O. Reinoso, and R. Aracil (Spain)


Robotics, Teleoperation, Grasping, Machine learning, Decision trees.


A robotic application is presented. Its purpose is to automate grasping processes in teleoperated cells. Given an object of an arbitrary shape, the grasping points have to be chosen in order to allow a reliable grip. A learning based approach is proposed, where the training examples are given by the user through teleoperation. Previous related works calculate the grasping points analytically from the geometry of the object. The present approach gives better results as it considers other aspects apart from object geometry. Such information is implicit in the training examples given by the user. Two -or three- jaw grippers are considered, so a pair -or a triplet- of contact points needs to be calculated. A two step learning process is carried out using decision trees : first, a set of valid contact points is obtained among all the contour points; then, all the pairs or triplets of valid points are checked to find the optimal pair or triplet. Both a simulation environment and an experimental setup have been developed. Results shown at the end of the paper validate the approach.

Important Links:

Go Back