Adaptive Grasping: Machine Learning Applied to Planning of Power Grasps

Robots used for tasks in hazardous environments frequency must cope with environmental uncertainty. The objects which must be manipulated are not well known in advance. Multifingered dexterous grippers are useful for grasping objects in uncertain environments. A fully enveloping grasp, also called a power grasp, is well suited for stable grasping in such unstructured environments. This paper describes the development of an adaptive power grasp planner. The system is based on a neural network which is trained on a variety of primitive shapes. When presented with a new object, the planner can generalize to achieve a secure grasp. Essential to the training of the neural network is the concept of grasp quality, which is determined from the convex hull of applied unit forces in wrench space. The mathematical framework and examples of grasp quality are presented, along with a description of the dexterous hand used for experimentation.