As of 29 Mar 2010, I was ranked 2,474th nationally in squash (619 in men's 3.5 skill level).
Matt's Research Interests
Humans understand how the world works, in a way
that enables them to manipulate the world very effectively. Robots lack
that understanding. My goal is to discover the nature of that understanding.
Part of my time goes toward exploring the fundamental mechanics of manipulation:
how one's actions affect the world. Part of it goes toward automatic
planning: given some goals, what actions would be appropriate. Below
are brief descriptions of a variety of projects to explore these issues.
(Most of the below is pretty old. Here are links to some more recent
Desktop Robotics (Robotic
Is it possible to build robots that take the desktop
as their task domain? They should be able to perceive the state of a
desktop, to navigate the desktop, and manipulate objects commonly found
on a desktop. I've begun a project to explore the desktop as a task
domain. (I've been working with Mike
Erdmann, Illah Nourbakhsh,
and Daniela Rus.) Our
first system is a mobile robot which uses its wheels for manipulation
as well as for locomotion. Imagine a small car planting its front wheels
on a piece of paper, and using the rear wheels to drive the robot and
the paper around. At the same time, if the front wheels are powered,
the robot could use them to manipulate the paper. Or even more challenging:
imagine the car rolling its front wheels onto a pencil, then rolling
the pencil around on the desk, sort of like a dung beetle.
A related development is some work on rapid prototyping
of mobile manipulators and mobile robots. The RC Mobipulator (see videos
immediately above) was built in half a day using hobby robotics techniques.
A terse set of instructions is available here.
The Palm Pilot Robot
Kit is the most interesting example. It is primarily the work of
Greg Reshko, collaborating with Illah Nourbakhsh and Garth Zeglin. The
Palm Pilot Robot Kit is a design for an easy-to-build, fully autonomous
robot controlled by a Palm handheld computer.
Dynamic Manipulation (Robot
Juggling, of a sort)
Robots typically use static and quasistatic methods
to interact with the world. People, on the other hand, are adept with
dynamic methods. Some scientists, notably Bill
Calvin, have argued that the evolution of the human brain was driven
by the challenges of accurate throwing. It is an interesting challenge
to model-based robotics to develop robots that can exploit the dynamics
of a task domain. Kevin Lynch's
PhD thesis demonstrated several instances---a snatch, a throw, and a
rolling throw. Each of them is planned automatically using information
about the object such as its shape and mass, and also with a good model
of the dynamic behavior of our arm. Here's a video of a rolling throw:
Usually we think of robots moving things around
by picking them up, but pushing is sometimes better. Think of rearranging
your furniture, gathering coins off your desk, or packing the trunk
of your car to get an idea of the advantages. As another example, note
that the early influential SRI robot SHAKEY worked by pushing things
around on the floor, as do many modern mobile robots.
My 1982 PhD thesis was about pushing. Mainly it was about the mechanics
of pushing--about which way a pushed object will move. It also addressed
the problem of automatic planning, of finding a pusher motion that will
move an object as desired. My thesis was published by MIT Press, along
with Ken Salisbury's thesis, under the title Robot Hands and the
Mechanics of Manipulation. But the best publication for this work
is the paper in the International Journal of Robotics Research.,
v5 n3, 1986.
In collaboration with Randy Brost, Ken
Akella, and Kevin Lynch,
I have continued to work on pushing. Randy Brost applied the work to
a grasping problem, and developed an automatic planner. Ken Goldberg
further refined the grasp planning ideas, and developed some potential
uses in automated manufacturing. Srinivas Akella developed and implemented
a planner for pushing a polygon to any specified position in the plane,
and proved that any polygon can be manipulated. Kevin Lynch worked on
several different aspects of pushing, including basic mechanics, model
estimation, and automatic planning.
is something to play with, an automatic push planning system that you
can access over the web.
Compliant Motion and Force
"Compliant motion" means programming
a robot to react gracefully when it comes into contact with other objects
in a task. Often compliant motion is achieved by controlling the force
applied by a robot, rather than controlling its motion. I worked on
compliant motion for my master's thesis, and for subsequent publications,
finishing around 1982. My main publication was a paper in theIEEE
Journal of Systems, Man, and Cybernetics in 1981, but for a quick
introduction to this work see John Craig's textbook. I also edited a
chapter and contributed to the introductory section of the 1981 MIT
Press collection Robot Motion: Planning and Control..
Factory Automation, Robotics
as a Natural Science
If you believe:
- factory robots do not use many sensors
- factory automation is uninteresting because
all the uncertainty has been engineered away
- factories mostly employ devices so simple that
they can scarcely be called robots at all
then you need to be enlightened! An automated factory
should be viewed as a system, comprising possibly hundreds of degrees
of freedom and also hundreds of sensors. Automated factories use every
trick in the book, and also some tricks not in the book, to deal with
uncertainty. If you are interested in how robots deal with uncertainty,
an automated factory is an excellent source of inspiration.
To study automated factories is to treat manipulation
as a natural science. Manipulation occurs all around us, and you can
learn a lot just by watching. A factory is a good place to watch. There
is lots of manipulation going on, and although it is highly evolved
it still uses mainly simple devices and techniques. Automated factories
are such great scientific assets that we have two specimen factory automation
systems in the Manipulation Lab. One is the SONY SMART Cell, which performs
a partial assembly of a SONY Walkman cassette player. It includes an
APOS system which automatically orients and feeds parts, and a robotic
manipulator with six different hands used to grasp and assemble the
parts. The other system is an ADEPT 550 manipulator along with a FlexFeeder
which uses a set of conveyors combined with a vision system to orient
Our most interesting result in factory automation is called 1JOC,
which stands for "One Joint Over Conveyor". This is joint
work with students Srinivas
Huang, and Kevin Lynch.
The idea is to manipulate objects on a conveyor belt, using a system
that is very simple but still programmable. We found a technique using
a very simple system--a single joint robot--and proved it is possible
to manipulate arbitrary planar shapes on the conveyor.
Model-Based Robots (Robots
that translate goals into actions)
All of the projects above reflect a basic interest
in model-based robots, and in reasoning about uncertainty.
A model of how the real world works gives a robot a way of trying
an action out in its head before taking the time and risk of trying
the same action in the real world. A model-based robot can plan elaborate
sequences of actions, achieving complex sophisticated goal-directed
No model can be perfectly accurate. Likewise, neither
motors nor sensors are perfectly accurate. There are lots of interesting
ideas about how to deal with uncertainty and error. Some models deal
explicitly with uncertainty, so that the robot can combine all sources
of information to reduce uncertainty, or choose actions which minimize
the effects of uncertainty and error. At the same time it is possible
to use machine learning techniques to refine the model on the basis
of previous experience.
last modified 5 Dec 2007.