杏吧原创

THE shelves of Mark Tilden鈥檚 lab are lined with hundreds of
robots鈥攔obots that walk, robots that roll, robots that crawl suicidally
through minefields, even robots designed to float in space. Most are small
enough to sit on the palm of your hand, and seem almost toy-like. The impression
is reinforced by the green rubber dinosaur and the model of the Star
Wars robot R2-D2 which share shelf space with them.

But inside these tiny machines something intriguing is going on. Using simple
networks of a few transistors, resistors and capacitors, and avoiding
conventional computer processors entirely, Tilden has been able to coax complex,
adaptive behaviours out of his robots. They are brainless wonders.

This approach is at odds with mainstream robotics, which has grown up with
digital technology. Whether it鈥檚 a new, custom-built industrial robot or the
latest automaton to display artificial intelligence (AI), most cutting-edge
machines rely on microprocessors.

Tilden, who works at the Los Alamos National Laboratory in New Mexico,
believes this approach is doomed to failure. He calls it the Asimovian approach,
after the science-fiction writer Isaac Asimov, who inspired the first generation
of modern robot designers. It鈥檚 based on the idea 鈥淚鈥檓 a rational being,
therefore I should be able to design another rational being,鈥� he says. The
trouble is that you have to increase processing power exponentially in order to
achieve a linear increase in ability. 鈥淭o make a robot that makes you toast
costs $20 000. To get that same robot to butter your toast costs an
additional $300 000.鈥�

He speaks from experience. Early on he tried to design a processor-based,
jack-of-all-trades house-cleaning robot. In the end the complexities of
programming the machine got the better of him. 鈥淣ow I use it as a hat rack,鈥� he
says. So Tilden has gone to the other extreme. His approach is to build cheap,
simple robots that are competent at a single task. If you want another task
performed, you design another simple robot.

In his home he has dozens of robots to do the cleaning. Some of them are
floor cleaners鈥攖hey constantly skitter around, sweeping the floors and
delivering the dust to a central depot. Each can clean only a little bit, but
since there are several of them, and all they do all day is clean floors, the
floors stay immaculate. Others do nothing but glide up and down his windows,
cleaning as they go. 鈥淚t鈥檚 a balanced robot environment,鈥� he says. 鈥淚t鈥檚 a robot
别肠辞濒辞驳测.鈥�

A matter of timing

All this and not a microprocessor in sight. So how do these creatures work?
At the heart of them all is what Tilden calls a nervous net鈥攆or which he
holds the patent. The net is essentially a ring of artificial neurons made up of
pulse delay circuits. Each circuit consists of a resistor, a capacitor and an
inverter, and stores up charge until it reaches a certain threshold. Then it
fires. By hooking up the output of one neuron to the input of the next, Tilden
can send a timed pulse round the ring that pauses at each neuron in turn before
moving on to the next.

The central controller for, say, a four-legged robot, would take the shape of
a ring of four neurons, one for each leg. Each neuron would in turn connect to a
further four 鈥渓eg neurons鈥� that command two motors鈥攐ne to move the leg up
and down, the other to manage motion forwards and backwards. A pulse passes
round the controller, but as each neuron fires it also causes the leg neurons to
fire, moving each leg in turn. For this 鈥渃reature鈥�, walking is not a carefully
programmed task, but something that simply emerges from the interplay of the
neurons.

Another clever aspect of the nervous net is that, so long as the motors take
their power from the same source as the control pulse, a simple form of feedback
is built into it. When a leg is slowed by an obstacle, the motor draws more
power, leaving less for pulse generation. This automatically slows down the
walking sequence until the leg is free to move again.

Unpredictable critters

In more complex applications, sensors can be added to enhance this feedback
effect. By injecting their own pulses into the ring, these sensors cause the
central control neurons to fire with different timings, modifying how the robot
moves. By properly tuning the timings, a pulse from a sensor which indicates
that the robot has hit an obstacle might make the legs move so they turn the
robot left or right to avoid the obstacle.

As their nervous nets become more complex, the robots鈥� behaviour becomes
harder to predict. Tilden designs the nets with particular behaviours in mind.
But then he has to tinker with the configuration, finding out through trial and
error which behaviours are produced by which adjustments.

To demonstrate his machines, Tilden puts a dozen in a big, shallow box and
carries them outside to a loading bay. The robots are all solar powered (most
have solar cells scavenged from old pocket calculators), and under the bright
New Mexico sun they perk up. They all have wheels, and their mission is to keep
moving. To do this, they head towards light sources and negotiate obstacles.
Each is surrounded by sensors鈥攚ires that form what look like all-round
bumpers鈥攚hich tell the robot when it has hit an obstacle.

When the obstacle is the side of the box, the robots just keep turning until
they鈥檙e clear. But when the obstacle is another moving robot, getting clear can
be more difficult. It鈥檚 in these robot scrums that different 鈥減ersonalities鈥�
become apparent from different nervous net configurations. Some robots avoid
fights and move away quickly. Others are more aggressive. They ram other robots,
repeatedly triggering their victims鈥� sensors and flooding them with a fast and
erratic flow of pulses. Eventually, the victims鈥� systems overload and freeze up
until a coherent series of pulses starts to flow again.

One robot is so aggressive it is self-destructive. It moves quickly,
banging repeatedly into any other robot it comes across. But it always ends up
overloading itself and stops in confusion. Tilden likens it to a guy in a bar
who doesn鈥檛 know when to stop picking fights.

But the nervous net approach is not limited to playing solar bumper cars.
Tilden and his colleagues have proposed launching hundreds of tiny 14-gram robot
satellites into space and using them to measure the magnetosphere, or equipping
each with a small imager and then combining images from the entire swarm to make
a single high-resolution picture
(New 杏吧原创, This Week, 23 March 1996).
To keep themselves oriented correctly, these 鈥渟atbots鈥� would use
electromagnets connected to a light detector via a simple nervous net. The light
detector would point at the Sun, and if it wandered from this position the robot
would power up the electromagnets to realign it by pushing off the Earth鈥檚
magnetic field.

In the lab, Tilden demonstrates a prototype that consists of three arms
arranged like a camera tripod. At the end of each arm is a coil of copper wire
that serves as an electromagnet. Tilden balances the robot on a stand and brings
a lamp closer. The satellite rocks as it tries to orient itself towards the
light. Gravity defeats it, but in space it would have no trouble.

In a more down-to-Earth application, Tilden has taken some of his walking
robots to the Yuma Proving Grounds in Arizona, to see if they could find and
dispose of landmines and unexploded ordnance. He reasoned that his machines have
a number of advantages for the work. For a start, they can travel across rough
terrain more easily than wheeled devices. The simplicity and robustness of their
circuitry also means his robots could work in conditions of heat, cold, mud or
dust that would defeat more complicated machines. And they are more likely to
survive an explosion. Even if two legs have been blown off, his robots can still
operate, dragging themselves to the next mine. Even if they are flipped over,
they can still crawl along upside down on their 鈥渒nees鈥�.

Robots don鈥檛 have to be geniuses to clear mines. One plan is to use numerous
cheap machines that would simply wander at random, stepping on mines and
exploding them. Or they could be fitted with sensors to detect the mines. They
would then either mark them for later retrieval, pick them up and carry them to
a central site, or simply blow them up.

Tilden says he ran into one unexpected problem. The soldiers who witnessed
the exercise, seeing a crippled robot continue to drag itself suicidally along
looking for more mines, felt sorry for the machine. This is one of the most
fascinating things about Tilden鈥檚 machines. His only design criterion is that
they are competent at their given task, yet in carrying out that task they
behave in such complex ways that observers ascribe personalities to
them鈥攅ven intelligence. But is that justified? 鈥淎t what point will
competence turn into intelligence? I don鈥檛 really care,鈥� Tilden says. 鈥淭oo many
people are working on intelligence out there.鈥�

Beyond the best brains

The notion that intelligence could arise from such simple machines is not as
crazy as it may seem. It has been championed throughout the 1990s by one of the
gurus of robotics, Rodney Brooks of the Massachusetts Institute of Technology.
He rejected the idea of a central brain, preferring instead to distribute
control to simple components. He showed how complex behaviours arise from the
interaction of those components. In his six-legged robot, for example, each leg
is controlled independently and walking emerges by timing the actions of the
legs. Brooks argues that intelligence also emerges from the way
things鈥攑eople and robots鈥� interact with the world.

Tilden admits to being heavily influenced by Brooks. But Brooks relies on
digital logic and microprocessors, and this is where Tilden parts company with
him. 鈥淲hen something gets beyond a certain degree of complexity, no mind on the
planet, no amount of work, will get around it,鈥� he says.

Tilden鈥檚 analysis of the problems of microprocessor-based robotics is overly
pessimistic, says Illah Nourbakhsh, an AI specialist and roboticist at Carnegie
Mellon University in Pittsburgh. Take Tilden鈥檚 assertion that processor-based
robots need exponentially more computing power to achieve linear increases in
ability. This is true in theory, says Nourbakhsh. But in practice robot
designers are getting better at figuring out algorithmic shortcuts, allowing the
robots to make decisions without having to work through near-infinite decision
trees. And advances in computing are also making a big impact, he says. 鈥淚t鈥檚
definitely the case that many of us in the robotics field do better and better
as computers get better.鈥�

Robot rescue

Up to now, it is mostly processor-based robots that have found practical
uses. Researchers at Carnegie Mellon, for example, sent a processor-based
walking robot, Dante, into an active volcano in Alaska in 1994. On the other
hand, as Tilden points out, Dante eventually slipped and fell, and had to be
rescued by helicopter.

Relying too much on computers has led to undesirable consequences, says
Tilden. Many robots are not truly autonomous but are 鈥減uppets鈥濃�攎echanical
bodies tethered by a data bus to a powerful workstation. Others exist only as
simulations within computers. 鈥淒o you know what it鈥檚 like to go to a robot
conference and not see one robot?鈥� says Tilden. The real challenge, he argues,
is to make robots that can deal with the chaos of the everyday world. 鈥淭he
inside of a computer is a perfect world,鈥� he says. 鈥淭he real world is fractal.
It鈥檚 complex. It鈥檚 dusty.鈥�

In 1989, in a bid to get more people building real robots, Tilden founded the
BEAM Robot Games (BEAM stands for Biology, Electronics, Aesthetics and
Mechanics). In games held all over the world, contestants come together to see
whose robot can walk, jump, swim or fly the best. Tilden hopes that as more
people try out their own approaches, the robots will evolve into more efficient,
more competent machines.

Because BEAM robots are cheap and relatively easy to make, a large community
of enthusiasts has sprung up. On the Internet dozens of sites offer plans and
tips for building robots out of old calculators and cassette recorders. One of
the best is run by someone who says he鈥檚 still in high school.

Yet if Tilden鈥檚 robots are so great, why aren鈥檛 they being used for real
applications? It鈥檚 partly a schizophrenic public attitude to robots in general,
says Tilden. People expect a robot to be a sort of mechanical man, and nervous
net robots are far from that. Some are also afraid of robots鈥攕omething he
calls 鈥淭erminator phobia鈥�. When his own mother visits, he has to box up all of
his house-cleaning robots because they make her nervous. 鈥淢y mom is convinced my
last words are going to be, 鈥楴o, no! Back!'鈥�

For his robots, there is also one specific limitation. So far, it鈥檚 been hard
to get high-level functioning out of his robots. While the nervous nets perform
many surprisingly complex behaviours, they鈥檙e all reflex. They will never be
able to learn, or conduct any sort of planning or decision making.

The obvious next step is to add some sort of brain to these
bodies鈥攑erhaps a head with a camera for an eye鈥攁nd let it guide the
action. His analogy is a rider and horse: the rider guides the horse, but the
horse decides where to put its hooves. A conventional microprocessor could serve
as the rider, but you can hear the distaste in Tilden鈥檚 voice at the idea.
Besides, he says, it鈥檚 been tried and hasn鈥檛 worked well: conventional digital
processors don鈥檛 mesh well with nervous nets.

Tilden would rather use a neural network for the rider. His idea is to wire
up his artificial neurons in a complicated array that resembles the mesh of
neurons in the human brain. Like a nervous net, it gives complex performance
with relatively simple circuits. Unlike a nervous net, it is capable of changing
over time鈥攊t can learn
(see 鈥淐ell Wars鈥�, New 杏吧原创, 21 February, p 36).
The nervous net would still be there, controlling the body. But
the neural net would do the processing, figuring out which direction to go and
what task it wanted to accomplish.

One of Tilden鈥檚 first walking robots used a neural network. But it didn鈥檛
work. He still has the machine around, and he fetches it. It starts to walk, but
when it runs into a difficult or confusing circumstance鈥攊n this case
Tilden pushing it off balance鈥攊t quickly becomes discouraged and, oddly,
sits down on its back end as if sulking. The interaction between neural net and
nervous net is too complicated, and tends to break down easily.

Tilden is working on a new design that he thinks will do a better job of
linking the neural and nervous nets. He hopes this will give rise to much more
complex, planned behaviours for his robots.

Nervous chess

By the summer, he hopes to have a robotic 鈥渃hess set鈥� that can play itself.
The game won鈥檛 be played a move at a time, but will be more like a structured
game of bumper cars. The pawns will be simple nervous-net robots battling enemy
pieces for possession of the board鈥檚 squares. But the more valuable
pieces鈥攓ueen, bishops and rooks鈥攚ill be capable of sensing what is
happening several squares away and planning their moves. The queen will be able
to direct pawns to areas where they are needed and, if she鈥檚 in trouble, will be
able to send a distress signal to rally her other pieces.

In the long run, Tilden鈥檚 ideas may gain ground. But in the short
term鈥攗nless you build one yourself鈥攜ou might have to wait a long
time to see a nervous-net robot. Today, the idea with the best prospects is the
satbot, which is being funded by the US Defense Advanced Research Projects
Agency. So it could be that Tilden鈥檚 robots will soon be more common in space
than they are on the ground.