The Economist - USA (2020-08-29)

(Antfer) #1

64 Science & technology The EconomistAugust 29th 2020


2

1

going to have millions of walking robots in
human environments,” says Jonathan
Hurst, co-founder of Agility Robotics, a
firm based in Albany, Oregon. It has just
launched Digit, a bipedal, two-armed robot
which has the look of an ostrich about it. At
present, Digit costs $250,000. But it is early
days. As more walking robots are put to
work their development will accelerate and
their production volumes increase, bring-
ing the cost of a machine like Digit down to
the tens of thousands of dollars.
This process is similar to the emergence
of flying drones. They once cost millions,
and had limited uses, until researchers
worked out how to make small aircraft hov-
er using multiple co-ordinated rotors.
These devices could fly easily and autono-
mously. Prices fell to $500 or less, and
multi-rotor drones are now employed for
all manner of jobs, from cinematography
to aerial surveying to delivering packages.
Some in the field of robotics think walking
robots have started down a similar path.
What changed? “We now understand
the mathematics of locomotion to a much
greater degree,” explains Dr Ames. Old-
school walking robots, such as Asimo, a
famously cheesy android unveiled in 2000
by Honda, a Japanese carmaker, have stilt-
ed gaits. They shuffle along, placing one
foot forward, checking their balance, mov-
ing the other foot, rechecking their bal-
ance, and so on. “When you are walking,
you don’t do that,” he says. “Your feet are
just coming down and catching yourself.”

Stroll on!
The way that humans walk is sometimes
described by biomechanists as controlled
falling. Making a stride involves swinging a
leg out and placing it down with small sub-
conscious corrections to maintain stability
as the mass of the body above it shifts for-
ward. Each leg works like a spring. These
movements are predictable, and in recent
years researchers have found out how to
model them mathematically. Together
with better actuators to operate a robot’s
limbs, and sensors which can measure
things more accurately, these models have
made it possible to recreate this style of
walking in robots. It does not require any
fancy machine learning or artificial intelli-
gence to do so, just good old-fashioned
computation, adds Dr Ames.
The difference between Asimo’s gait
and that of the new breed of bot is striking.
Whereas Asimo’s chunky legs look leaden,
Digit strides confidently along on a lean
pair of limbs, happily swinging its arms as
it goes. Atlas (pictured), an experimental
humanoid made by Boston Dynamics, is
more capable still. It can walk, run, jump
and even perform backflips. Asimo did a lot
of celebrity photo-opps, but it never went
into production. Honda quietly stopped
work on the project in 2018, to concentrate

on more “practical” forms of robotics, such
as mobility devices for the elderly.
It is easy to conclude, as many do, that
these new walking robots simply mimic
nature. But that is not quite the case. A
quadruped, being a stable platform, is a
good starting point from which to design a
walking robot. After co-ordinating the four
limbs, getting a good balance and fitting a
system of vision that lets the robot work
out where to put its feet, Spot’s designers
ended up with a dog-oid. Michael Perry,
head of business development for Boston
Dynamics, says that is not surprising be-
cause nature has been developing efficient
designs for a long time.
Another example of art evolving to imi-
tate nature occurred during the design of
Digit. This inherited its ostrich looks from
Cassie, a two-legged torso which Agility
sold to a number of research groups. Cas-
sie’s developers had to find a way to stop
some of the robot’s actuator motors from
working against each other. Their solution
turned out to look like a pair of bird’s legs.

Cassie subsequently acquired arms and
evolved into Digit as the result of the engi-
neers’ attempts to solve another problem.
When it swung a leg forward Cassie’s body
twisted a little, which sometimes caused
the robot to fall over if it was walking quick-
ly. In nature, some animals use tails to im-
prove their balance when manoeuvring at
speed. Borrowing this idea, Agility’s re-
searchers attached a pair of tail-like appen-
dages, one on each side of the robot’s torso,
to improve its mobility. That worked. Then
they turned the appendages into a pair of
arms. These can catch the robot should it
fall, and help it get up again.
The arms can perform other useful
tasks, too, such as moving boxes in a ware-
house. Digit can carry up to 20kg. Distribut-
ing and delivering goods is likely to be an
important application for walking robots,
reckons Dr Hurst, especially now that e-
commerce has boomed as a result of re-
strictions imposed in the wake of covid-19.

Some automated distribution centres are
set up for conventional fixed and wheeled
robotic systems, but these have usually
been built this way from scratch. Most
warehouses are designed around people.
Robots with legs, which move in a similar
way to human workers, would fit right in.
With further development, walking ro-
bots will undertake more complex tasks,
such as home deliveries. Ford is working
on this with a Digit robot that rides in the
back of a van. Though robots with wheels
already make some deliveries, reaching
many homes is tricky, and may involve
climbing steps or stairs. “Legs are how you
would want to get up to most front doors to
deliver a package,” observes Dr Hurst.
Exactly how this might be done remains
to be seen. Unless they are on a pre-
programmed mission, most mobile robots
require an operator to provide basic in-
structions to, say, proceed to a certain
point. The robot then walks there by itself,
avoiding obstacles and climbing or de-
scending steps and stairs along the way.
This means a walking robot making door-
to-door deliveries might need some kind of
digital map of the neighbourhood, to know
in advance the paths it can traverse and the
flower beds it should avoid. That might in-
volve a big data-acquisition effort, much
like those used to build digital maps for
driverless cars. Similarly, in a factory or a
warehouse, a walking robot would need to
be shown the ropes by a human being be-
fore it was let loose to work on its own.

Self determination
A fully autonomous robot that could walk
into an unknown environment and decide
for itself what it needed to do remains a
long way off. One of the hardest tasks for
such a device would be caring autono-
mously for someone at home. The robot
would have to be able to make numerous
complex decisions, such as administering
the correct medicine, deciding whether or
not to let strangers into the house or know-
ing when to take the dog for a walk. Yet
many roboticists think they will get there,
or at least close to it, one day.
In the meantime, the new generation of
robots now being developed will keep
building up the machines’ capabilities. At
Boston Dynamics Mr Perry reckons that,
besides surveying, Spot will find many
roles in inspection and maintenance. Such
robots can, for instance, enter hazardous
environments like electrical substations
without them having to be taken off the
grid, as is necessary whenever a human en-
gineer goes inside.
Instead of just looking for problems,
Spot’s next trick will be to take action to re-
solve them, such as throwing a switch or
turning a valve. It will do this with a single
manipulator arm which makes it look less
like a dog and more like a long-necked
Free download pdf