Matthew Mason

Professor

Office: A521 Newell-Simon Hall

Email: matt.mason@cs.cmu.edu

Department:

Robotics Institute

Computer Science Department

Web site:

http://www.cs.cmu.edu/~mason/

I work on robotic manipulation. The ability to deal with almost any random object that comes along is the most astonishing thing that humans and other animals do, and the most intriguing challenge in robotics.

My main focus is the Simple Hands Project. Here's the idea. Most robot hands can be classified as either simple or complex. Simple hands are like pliers or tongs, but highly specialized. Often they are designed to handle just one specific object. Despite this limitation, simple hands are by far the most common in practical applications.

Because of the highly specialized nature of most simple grippers, many roboticists work on complex hands, usually anthropomorphic, with lots of fingers, joints, motors and sensors. Despite decades of work on these hands, their success in general-purpose manipulation is still only a distant dream.

Here's an interesting observation on simple hands. When attached to a brain, rather than a computer, they are very capable. A human with a simple prosthetic hook can run circles around any robotic system. There are two lessons to learn from this. The first lesson: simple hands do not have to be specialized. The second lesson: general purpose manipulation is a function of the brain, not the hand. This suggests that robotics research should shift focus from complex hands to simple hands. Simple hands are easier to understand, and hence better suited to scientific research. Of course they are also tougher, lighter, cheaper ... they are handier.

Simple Hands is primarily a computer science project, but with important contributions from applied mechanics. The main challenges are perception and planning. Perception uses all available information to determine what is in the hand, and planning chooses actions to achieve an end goal. Both are founded on a machine learning approach, employing experimental data from thousands of trials. In short, the robot messes with things to learn how they behave, how to sense them, and how to manipulate them.