Machine Learning and Friends Lunch
From Supervisory Control To Intelligent Prosthetics
Michael Rosenstein
UMass
Abstract
The numbers are staggering. In the United States, more 30,000 people have lost an arm. More than 25,000 diabetes-related amputations could be prevented each year with proper foot care. And more than 80,000 yearly occurrences of traumatic brain injury result in long-term disability. Prevention is essential for making a dent in such figures, but for individuals living with disease and disability, technology--prosthetic devices in particular--may hold the key to dramatic improvements in quality of life. My goal for this talk is to show that supervisory control and machine learning offer a novel approach for developing intelligent prostheses.
With supervisory control, a human operator intermittently takes control of a process that is otherwise controlled by a computer. The first part of the talk will cover my past robotics work where the human operator is replaced by a stable controller that supervises a machine learning system. The task of the learning system is to control a robotic arm in an optimal fashion, and the stable, yet sub-optimal controller essentially speeds up learning while constraining the admissible actions to those that are safe. The second part of the talk will cover my recent work on human-robot interaction. In particular, I'll describe a user interface that predicts operator intentions and that allows a robot to respond appropriately to commands but without the fine-grained, fatiguing input common to remotely operated systems. In the final part of the talk I'll describe how both lines of research hold potential for improved control of artificial limbs.