Grasp Planning Using Low Dimensional Subspaces

In this chapter we explore the concept of low-dimensional posture subspaces for artificial hands. Recent advances in neuroscience research have shown that control of the human hand during grasping is dominated by movement in a configuration space of highly reduced dimensionality. This has led our group to explore how artificial hands may take advantage of similar subspaces to perform complex grasping tasks. Subspaces are important not only because they are biologically motivated but because they allow us to create computational frameworks that are tractable for difficult problems, characterized by a large number of degrees of freedom, such as dexterous grasping. The work described in this chapter allows us to compute metrics on thousands of grasps quite efficiently, leading to a sampling-based approach that can adequately cover the space of grasps for a number of robotic hands as well as a human hand model. This approach is based on a hand posture subspace defined by a small number of basis vectors which we call eigengrasps. The implied dimensionality reduction has allowed us to perform online dexterous grasp planning both for robots needing to find a correct grasp for an object and for prosthetic devices in which the human provide a subset of the necessary Degrees of Freedom (DOFs), allowing the planner to work in real-time to find a stable grasp. Further, the ability to pre-compute thousands and thousands of stable grasps for dexterous hands over a large class of objects has motivated a new direction in grasp synthesis, which we call data driven grasping. If the number of objects to be grasped in the database is very large and comprehensive then robotic grasping becomes a pre-computed indexed database lookup which is extensible to grasping novel objects not in the database.