"Data-driven grasping and manipulation"

Nancy Pollard, Carnegie Mellon University

Data captured from human performances of activities ranging from the everyday through the extraordinary has become widely accessible over the past 10 years. The ability to download or capture human motion and process it in real-time has led to many new algorithms and new ways of thinking about character animation and robot control. However, we do not yet know how to make the most effective use of this data. What is important about a given performance? How can it be modified to create realistic new scenarios? And what are the limits of this approach.can we ever create behavior that could be called dexterous from a collection of observed performances?

In this talk, I will focus on the problem of creating dexterous grasping and manipulation behaviors from observed performances. I will discuss how my ideas have changed over the past decade, as we have gone from the idea that a grasp is made up of contact points between the hand and object through consideration of the hand geometry, anatomical constraints, and dynamic properties to the observation that grasps often involve preparatory sensing and manipulation actions which we have shown can reduce the effort needed to acquire an object. Results in computer animation and robot control, as well as results from controlled human subjects experiments will be presented.