RR:Collaborative Research Resources: Learning from Human Hands to Control Dexterous Robot Hands

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

This project, investigating human grasping and manipulation strategies to gain insight into synergies employed by human subjects in grasping and manipulation tasks, aims to explore how best to use these findings to create better control algorithms for robot hands. Of particular importance are techniques to make autonomous robot behavior robust to uncertainties and to make algorithms with a human in the loop, such as teleoperation and control of prosthetic devices, more intuitive and effective for fine manipulation tasks. Although a high degree of freedom (DOF) device may be required to manipulate a wide variety of objects and perform a wide variety of tasks, in any given task situation, only a small number of independently controlled DOF may be necessary. To make significant progress towards dexterous grasping and manipulation, we must: Make the best possible use of available sensing technology (specially for force sensing), and Understand how to analyze, plan, and control hand motion in a reduced degree of freedom space (take advantage of task-based coordination rules and synergies that may make real time grasp optimization and planning tractable). The infrastructure should contribute in answering the following questions: What is an optimal grasp, and how does it depend on the kinematic and dynamic properties of the device doing the grasping? What is the relationship between critical signals, muscle activation levels, hand shape, and force production during task performance? Does analysis of data collected along the pipeline from cortical signals to force production allow easy organization into grasp primitives or result in "control handles" that a human operator of a robot hand would find intuitive? How can human demonstrations of grasping and manipulation tasks be employed as the wonderful resource they seem to be, i.e., how can individual examples be converted into control algorithms that will function on a robot and be robust to variations and uncertainties? Broader Impact: The results will improve the understanding of human hand motion and force production, the use of force information in teleoperation and control of prosthetic devices, and the ability to coach robot behavior through task demonstration. Teams of undergraduate students will use the facility and data collected will be made available on the web.