[info] wired: more on robots learning to grab things

[Alejandro Dubrovsky]

(
and other semi-related robot projects
http://www.wired.com/science/discoveries/magazine/15-12/mf_robothand?currentPage=all
)

Getting a Grip: Building the Ultimate Robotic Hand
By Gregory Mone Email 11.27.07 | 12:00 AM
UMan uses trial and error to figure out how to manipulate items it has
never seen before.
Photo: Glenn Matsumura

A 6-foot-tall, one-armed robot named Stair 1.0 balances on a modified
Segway platform in the doorway of a Stanford University conference room.
It has an arm, cameras and laser scanners for eyes, and a tangle of
electrical intestines stuffed into its base. It's not pretty, but that's
not the point. From his seat at a polished table, roboticist Morgan
Quigley sends the bot on a mission. "Stair, please fetch the stapler
from the lab."

Nothing happens. Quigley asks again. Nothing. After the third attempt,
Stair responds in an inflectionless voice: "I will go fetch the stapler
for you."

Using its laser scanners to identify potential obstacles, Stair 1.0
rolls out of the room and into the lab's central workspace, a
rectangular area bordered by desks. On one side is a kind of robotic
graveyard, a jumble of decades-old industrial arms. A poster of the NS-5
humanoid from the movie I, Robot seems to taunt the researchers from its
spot on the wall: Try building me, punks. Quigley and computer scientist
Andrew Ng, who directs the Stanford AI Robot (Stair) project, walk
behind their robot, watching.

Stair 1.0 searches the rows of workstations, then locates the stapler.
The robot moves forward and stops. If it had lungs, it might take a deep
breath, because this is the hard part.

Up to this point, Stair hasn't done anything all that impressive. Plenty
of robots can move around a room — or, as the Darpa Grand Challenge
unmanned vehicle race proved, navigate far-more-complex terrain, like
the open desert. But now Stair is going to switch from observing and
navigating the world to interacting with it. Instead of just avoiding
obstacles, the robot is actually going to manipulate something in its
environment.

Yes, robots already play the trumpet, sort chemicals in labs, weld cars.
But these bots are just following a script. Shift the pieces along an
assembly line and the robot won't be able to build a bucket, let alone a
Buick. And outside those controlled environments, objects and people
don't stay put. Staplers are misplaced. Scripts don't apply.

Yet Stair 1.0 seems to be doing fine. It locates the stapler and
stretches out its hand, a simple, two-fingered gripper with foam padding
taped on to serve as makeshift skin. Three minutes after Quigley spoke
his initial request, the robot reaches down, closes its fingers, and
lifts its hand up from the table.

And all it holds is a pocket of air.

To do real work in our offices and homes, to fetch our staplers or clean
up our rooms, robots are going to have to master their hands. They'll
need the kind of "hand-eye" coordination that enables them to identify
targets, guide their mechanical mitts toward them, and then manipulate
the objects deftly.

There's a growing need for robots with these skills. In Japan, the elder
care industry is already employing robots as assistants. To keep seniors
out of costly nursing homes, though, they need to be able to perform
household chores like serving up a drink. Even that simple task will
entail plucking a glass out of a crowded cupboard, finding and removing
a bottle from a fridge, and then pouring the beverage from one container
into the other. And the bot needs to do all this without spilling,
dropping, or breaking anything.

These helpful machines don't have to be perfect, though. Occasionally, a
glass will fall. Robots will have to be programmed to fail gracefully,
and, more important, to learn from those failures. That's where Stair
1.0 came up short. In going for that elusive stapler, the bot did
everything right — until it failed to notice that it wasn't holding
anything. But the next generation, Stair 2.0, will actually analyze its
own actions. The next Stair will look for the object in its hand and
measure the force its fingers are applying to determine whether it's
holding anything. It will plan an action, execute it, and observe the
result, completing a feedback loop. And it will keep going through the
loop until it succeeds at its task. It sounds like a sensible enough
approach, as long as scientists can, in just a decade or so, engineer
the coordination and dexterity that evolution took millions of years to
perfect. The trick is to build robots that act more like children than
machines.

When a computer fails at a task, it spouts an error message. Babies, on
the other hand, just try again a different way, exploring the world by
grabbing new objects — shoving them into their mouths if possible — to
acquire additional data. This built-in drive to explore teaches us how
to use our brains and bodies. Now a number of hand-focused roboticists
are building machines with the same childlike motivation to explore,
fail, and learn through their hands. Stair and a robot called UMan at
the University of Massachusetts Amherst, two of the first robots
conceived from the hand up, will both get a mild version of a
kick-the-chick-out-of-the-nest education. Their creators plan to let the
robots learn through trial and error. Meanwhile, on the other side of
the Atlantic, a 4-foot-tall Italian humanoid is getting prepped for a
different — and completely unique — kind of schooling: It will learn
through imitation.

Barely past its second birthday, Stair 1.0 is already obsolete. The
upgrade, Stair 2.0, has the same basic home-built appearance, but it's
outfitted with a far more advanced hand, manufactured by Barrett
Technology in Cambridge, Massachusetts. The size of a catcher's mitt,
the BarrettHand has three oversize fingers. Two of them rotate around
the palm, switching positions, effectively giving the hand a pair of
opposable thumbs.

As the motionless Stair 1.0 sits in a cornersof the Stanford lab, PhD
student Ashutosh Saxena is getting Stair 2.0 ready for a test of its
skills. He moves Stair 2.0's arm around like a physical therapist, then
asks it to go to a dishwasher set up on the far wall.

Saxena instructs it to remove a cup from the rack, but he hasn't told
Stair how to do so. Instead, he and the other development team members
have equipped Stair with a set of algorithms that allow it to learn on
its own. One governs the bot's ability to identify an object in a loaded
dishwasher, another suggests the best way to move its hand toward that
object, and the third decides how to pick the thing up.

While Saxena watches, Stair tries several times to grab the cup. It
fails each time, but it records those actions as unsuccessful so it
won't repeat them.

Still, it's hard to watch, because to us the task looks so easy. The
robot should just move its hand directly over the cup, grasp it, and
then pull it up. "That's how I'd do it," Saxena must think.

Then Stair surprises him. Instead of taking the direct route, the robot
reaches around and repositions its arm so that it can move its hand
across the top rack, approaching the cup from the side. This time it
succeeds, and Saxena laughs. "It's funny to see the robot find its own
way," he says.

Funny, but also impressive: It shows that the robot is learning.

In a more spacious lab at the University of Massachusetts, UMan is going
through a similar kind of basic training. Stair and UMan could be
brothers: They look alike, use the same scanning lasers, and were both
developed around a single hand built by Barrett.

The UMan creators designed an algorithm that helps their robot figure
out how to use that hand with objects it has never seen before. To test
it, they built some toys for the machine-child, one of which is just
three long wooden blocks joined by two hinges, with a fourth piece that
slides in and out of one of the blocks at one end, like a drawer.

Because UMan has been programmed to experiment, to try things out, the
roboticists simply put the toy on a table in front of it and wait. After
UMan discerns the difference between the toy and the background — a
standard computer vision trick — the algorithm stipples the robot's
mental picture of the object with a series of points. Then UMan reaches
out, pushes and prods, and tracks the toy's movements by measuring how
the distances between all those points change. In doing so, it discovers
the location of all the joints and, in effect, how to play with the toy.

Using this same algorithm, the robot has already learned how to turn an
unfamiliar door handle or knob — something other machines have trouble
with. UMan mentally separates the handle from the door, pushes and turns
until it figures out how the handle works, then stores that experience
for future reference. Eventually, project leader Oliver Brock hopes, a
set of algorithms will allow his robot to accomplish more-complex tasks
— even things he didn't anticipate or build in at the start. "Human
babies spend a long time improving their manual skills," Brock says.
"Then they use those skills to learn new ones, like painting a window
frame or mowing a lawn."

But babies don't just wander around alone, picking up strange objects
and trying to figure out how they move — there wouldn't be a lot of
adults if that's how we spent our childhoods. Babies rely heavily on
others to show them what to handle and how to handle it. Some scientists
believe that this flavor of dependency is actually the key to robotic
independence.
RobotCub is shaped like a human so it can learn by imitating its
scientist "parents."
Photo: Glenn Matsumura

The grand green Apennine Mountains fill the windows at the University of
Genoa's Laboratory for Integrated Advanced Robotics, but otherwise it
isn't that different from the other labs: As Europe's preeminent
robotics facility and one of the world's epicenters of artificial
intelligence research, it's dominated by eggheads staring at monitors.
And, of course, there's an android hanging around the place.

The size and shape of a 3-year-old, RobotCub has two five-fingered
hands, each of which will be covered with sensitive artificial skin made
of the same stuff as the iPod's electrostatic touchwheel. It has
expressive eyes, a white plastic shell that makes it look like Casper
the Friendly Ghost, and a tether that runs from its back like an
electronic umbilical cord into an adjacent room, where it connects to a
few dozen PCs. These machines will be charged with running each of
RobotCub's 53 electric motors. They'll process the sensory information
it gathers through its hands and cameras and decide how to move the
machine in response. RobotCub might be the size of a child, but its
brain fills an entire room.

The experiments, due to start early next year, will seem simple. There
will be blocks on a table; Giorgio Metta, the lead roboticist on the
project, will take one of them and stack it atop another. Ideally,
RobotCub will study his action and, up in its processors, substitute its
own arms for Metta's, its artificial hand for his real one. Ideally, it
will then reinterpret what it witnesses and repeat the action using its
own hands. "This is where the robot's form is critical," Metta says.

RobotCub's humanoid shape and five-fingered hands are more than a dreamy
attempt to build an android. The tricky part about learning through
imitation is that the student has to have the same parts as the teacher.
That's why this method might not work with Stair or UMan. If Saxena had
pushed Stair aside while it was trying and failing to grab that cup out
of the dishwasher, if he had followed the father-to-son,
let-me-show-you-how-to-do-it method of instruction, his robot would have
been stumped. Stair has one arm, a single three-fingered hand, and looks
more like a moving appliance cabinet than Homo habilis.

But RobotCub has the basic physical characteristics of a human — a head
with two eyes, a body, two arms and two legs, two five-fingered hands.
Metta's group designed RobotCub this way so they could model its
cognitive architecture on what are called mirror neurons. Discovered by
Luciano Fadiga, one of the team's neurophysiologists, mirror neurons
help explain how we learn through observation: When we watch someone
swing a golf club, for example, the neurons in charge of kick-starting
that swing also fire in our heads, even if we're just sitting on the
couch. Fadiga coauthored the first paper to describe the phenomenon, and
now he's helping to integrate the principle into lines of code that
represent neurons in RobotCub's brain.

Before mimicking the block-stacking, RobotCub will need to experience
all the individual actions required — reaching, grasping, lifting — for
itself. When Metta starts to go for that block, RobotCub takes a series
of quick snapshots and, by tracking the progress of his "father's" hand
from one photo to the next, extrapolates after only 200 milliseconds
what Metta is doing. The robot guesses that Metta is reaching, and it
connects this to its own experience with reaching. Next, it guesses
which objects Metta is most likely trying to grab; it determines whether
it recognizes them and whether it knows how to pick them up. At every
step, it watches Metta, connects its observations to its own experience,
and, once the roboticist is finished, tries to string the movements
together just as Metta did. RobotCub should be able to learn how to
accomplish the same end — stack the blocks — in its own way. It should
be able to think, "OK, if I drive these motors like this and position
myself like so, I can put this block on top of that one, too."

It should be able to learn by watching.

Meanwhile, UMan is getting ready to learn by doing. Its next activity
will be to roll through the lab and open random doors, surprising
unsuspecting academics at their desks. And Stair 2.0 should soon be able
to find, heat, and serve that holy staple of the grad student's diet:
the frozen burrito. Whether any of these machines will be truly
intelligent is another issue. Building robots that work with their hands
is not about synthesizing Descartes. It's about getting machines to a
point where they can provide real value in our unstructured,
unpredictable world — be that assisting the elderly, cooking meals, or
doing the dishes. And just as our nimble hands got us into the
flint-and-fire game, this approach to robot development may be the spark
that gets these machines off the assembly line and into our lives.

Gregory Mone (gmmone at gmail.com), a writer living in Boston, wrote the
novel The Wages of Genius.