"Fast, cheap, and out of control" began appearing on
buttons of engineers at conferences and eventually made it to the title
of Rodney Brooks's provocative paper. The new logic offered a completely
different view of machines. There is no center of control among the
mobots. Their identity was spread over time and space, the way a nation
is spread over history and land. Make lots of them; don't treat them so
Rodney Brooks grew up in
Australia, where like a lot of boys round the world, he read science
fiction books and built toy robots. He developed a Downunder perspective
on things, wanting to turn views on their heads. Brooks followed up on
his robot fantasies by hopscotching around the prime robot labs in the
U.S., before landing a permanent job as director of mobile robots at
began an ambitious graduate program to build a robot that would be more
insect than dinosaur. "Allen" was the first robot Brooks built. It kept
its brains on a nearby desktop, because that's what all robot makers did
at the time in order to have a brain worth keeping. The multiple cables
leading to the brain box from Allen's bodily senses of video, sonar, and
tactile were a neverending source of frustration for Brooks and crew.
There was so much electronic background interference generated on the
cables that Brooks burnt out a long string of undergraduate engineering
students attempting to clear the problem. They checked every known
communication media, including ham radio, police walkie-talkies and
cellular phones, as alternatives, but all failed to find a static-free
connection for such diverse signals. Eventually the undergraduates and
Brooks vowed that on their next project they would incorporate the
brains inside a robot -- where no significant wiring would be needed -- no
matter how tiny the brains might have to be.
They were thus forced to use very primitive logic steps, and very
short and primitive connections in "Tom" and "Jerry," the next two
robots they built. But to their amazement they found that the dumb way
their onboard neural circuit was organized worked far better than a
brain in getting simple things done. When Brooks reexamined the
abandoned Allen in light of their modest success with dumb neurons, he
recalled that "it turned out that in Allen's brain, there really was not
The success of this profitable downsizing sent Brooks on a quest to
see how dumb he could make a robot and still have it do something
useful. He ended up with a type of reflex-based intelligence, and robots
as dumb as ants. But they were as interesting as ants, too.
Brooks's ideas gelled in a
cockroachlike contraption the size of a football called "Genghis."
Brooks had pushed his downsizing to an extreme. Genghis had six legs but
no "brain" at all. All of its 12 motors and 21 sensors were distributed
in a decomposable network without a centralized controller. Yet the
interaction of these 12 muscles and 21 sensors yielded an amazingly
complex and lifelike behavior.
Each of Genghis's six tiny legs worked on its own, independent of
the others. Each leg had its own ganglion of neural cells -- a tiny
microprocessor -- that controlled the leg's actions. Each leg thought for
itself! Walking for Genghis then became a group project with at least
six small minds at work. Other small semiminds within its body
coordinated communication between the legs. Entomologists say this is
how ants and real cockroaches cope -- they have neurons in their legs that
do the leg's thinking.
In the mobot Genghis, walking emerges out of the collective behavior
of the 12 motors. Two motors at each leg lift, or not, depending on what
the other legs around them are doing. If they activate in the right
sequence -- Okay, hup! One, three, six, two, five, four! -- walking "happens."
No one place in the contraption governs walking. Without a smart
central controller, control can trickle up from the bottom. Brooks
called it "bottom-up control." Bottom-up walking. Bottom-up smartness.
If you snip off one leg of a cockroach, it will shift gaits with the
other five without losing a stride. The shift is not learned; it is an
immediate self-reorganization. If you disable one leg of Genghis, the
other legs organize walking around the five that work. They find a new
gait as easily as the cockroach.
In one of his papers, Rod Brooks first laid out his instructions on
how to make a creature walk without knowing how:
There is no central controller which directs the body where to put each
foot or how high to lift a leg should there be an obstacle ahead.
Instead, each leg is granted a few simple behaviors and each
independently knows what to do under various circumstances. For
instance, two basic behaviors can be thought of as "If I'm a leg and I'm
up, put myself down, " or "If I'm a leg and I'm forward, put the other
five legs back a little." These processes exist independently, run at
all times, and fire whenever the sensory preconditions are true. To
create walking then, there just needs to be a sequencing of lifting legs
(this is the only instance where any central control is evident). As
soon as a leg is raised it automatically swings itself forward, and also
down. But the act of swinging forward triggers all the other legs to
move back a little. Since those legs happen to be touching the ground,
the body moves forward.
Once the beast can walk on a flat smooth floor without tripping,
other behaviors can be added to improve the walk. For Genghis to get up
and over a mound of phone books on the floor, it needs a pair of sensing
whiskers to send information from the floor to the first set of legs. A
signal from a whisker can suppress a motor's action. The rule might be,
"If you feel something, I'll stop; if you don't, I'll keep going."
While Genghis learns to climb over an obstacle, the foundational
walking routine is never fiddled with. This is a universal biological
principle that Brooks helped illuminate -- a law of god: When something
works, don't mess with it; build on top of it. In natural systems,
improvements are "pasted" over an existing debugged system. The original
layer continues to operate without even being (or needing to be) aware
that it has another layer above it.
When friends give you directions on how to get to their house, they
don't tell you to "avoid hitting other cars" even though you must
absolutely follow this instruction. They don't need to communicate the
goals of lower operating levels because that work is done smoothly by a
well-practiced steering skill. Instead, the directions to their house
all pertain to high-level activities like navigating through a town.
Animals learn (in evolutionary time) in a similar manner. As do
Brooks's mobots. His machines learn to move through a complicated world
by building up a hierarchy of behaviors, somewhat in this order:
Avoid contact with objects
Explore the world
Build an internal map
Notice changes in the environment
Formulate travel plans
Anticipate and modify plans accordingly
The Wander-Aimlessly Department doesn't give a hoot about obstacles,
since the Avoidance Department takes such good care of that.
The grad students in Brooks's mobot lab built what they cheerfully
called "The Collection Machine" -- a mobot scavenger that collected empty
soda cans in their lab offices at night. The Wander-Aimlessly Department
of the Collection Machine kept the mobot wandering drunkenly through all
the rooms; the Avoidance Department kept it from colliding with the
furniture while it wandered aimlessly.
The Collection Machine roamed all night long until its video camera
spotted the shape of a soda can on a desk. This signal triggered the
wheels of the mobot and propelled it to right in front of the can.
Rather than wait for a message from a central brain (which the mobot did
not have), the arm of the robot "learned" where it was from the
environment. The arm was wired so that it would "look" at its wheels. If
it said, "Gee, my wheels aren't turning," then it knew, "I must be in
front of a soda can." Then the arm reached out to pick up the can. If
the can was heavier than an empty can, it left it on the desk; if it was
light, it took it. With a can in hand the scavenger wandered aimlessly
(not bumping into furniture or walls because of the avoidance
department) until it ran across the recycle station. Then it would stop
its wheels in front of it. The dumb arm would "look" at its hand to see
if it was holding a can; if it was it would drop it. If it wasn't, it
would begin randomly wandering again through offices until it spotted
That crazy hit-or-miss system based on random chance encounters was
one heck of an inefficient way to run a recycling program. But night
after night when little else was going on, this very stupid but very
reliable system amassed a great collection of aluminum.
The lab could grow the Collection Machine into something more
complex by adding new behaviors over the old ones that worked. In this
way complexity can be accrued by incremental additions, rather than
basic revisions. The lowest levels of activities are not messed with.
Once the wander-aimlessly module was debugged and working flawlessly, it
was never altered. Even if wander-aimlessly should get in the way of
some new higher behavior, the proven rule was suppressed, rather than
deleted. Code was never altered, just ignored. How bureaucratic! How
Furthermore, all parts (departments, agencies, rules, behaviors)
worked -- and worked flawlessly -- as stand-alones. Avoidance worked whether
or not Reach-For-Can was on. Reach-For-Can worked whether or not
Avoidance was on. The frog's legs jumped even when removed from the
circuits of its head.
The distributed control layout for robots that Brooks devised came
to be known as "subsumption architecture" because the higher level of
behaviors subsumed the roles of lower levels of behaviors when they
wished to take control.
If a nation were a machine, here's how you could build it using
You start with towns. You get a town's logistics ironed out: basic
stuff like streets, plumbing, lights, and law. Once you have a bunch of
towns working reliably, you make a county. You keep the towns going
while adding a layer of complexity that will take care of courts, jails,
and schools in a whole district of towns. If the county apparatus were
to disappear, the towns would still continue. Take a bunch of counties
and add the layer of states. States collect taxes and subsume many of
the responsibilities of governing from the county. Without states, the
towns would continue, although perhaps not as effectively or as
complexly. Once you have a bunch of states, you can add a federal
government. The federal layer subsumes some of the activities of the
states, by setting their limits, and organizing work above the state
level. If the feds went away the thousands of local towns would still
continue to do their local jobs -- streets, plumbing and lights. But the
work of towns subsumed by states and finally subsumed by a nation is
made more powerful. That is, towns organized by this subsumption
architecture can build, educate, rule, and prosper far more than they
could individually. The federal structure of the U.S. government is
therefore a subsumption architecture.