91av

IN AUGUST an 800-kilogram robot called Dante picked its way into the crater
of Mount Spurr, an active volcano in southern Alaska. The crater was too
dangerous for humans to enter, and the terrain too rough for a wheeled vehicle
to negotiate. So Dante was fitted with eight legs. It walked around for a
couple of days, collecting air samples and transmitting video images back to
its human controllers, before it lost its footing and tumbled to the bottom of
the crater. Despite Dante’s last-minute stumble, the mission showed that
legged robots can work in places where even a Range Rover would not dare
venture.

For the past 20 years or more, researchers have been tinkering with similar
creations, inspired by the simple observation that legs work pretty well for
animals in rough terrain. Legged robots have been used to clean up nuclear
power plants, and are being considered for jobs as routine as maintaining
pipelines and as adventurous as exploring the surface of Mars. Like Dante,
many of these robots are bug-shaped – with plenty of legs to stand on they are
less likely to fall over. But previous generations of walking robots have not
been particularly graceful movers. Compared with the average cockroach,
walking robots have been a bunch of pitiful shufflers.

Now robot builders are going a step further. They are learning from the way
insects control their leg movements, and programming their robots to mimic
these mechanisms. Preliminary results are impressive, and a little eerie.
Rather than clunking along mechanically, the latest insect-like robots are
almost lifelike: they probe ahead with a leg to find a foothold, step casually
over obstructions, and pull back gracefully to regain their balance when the
floor is pulled out from under them. Roboticists hope eventually to build
walking robots that are as adaptable, robust and speedy as the common
cockroach, which can travel two to five times its own body length in a
second.

“The more I get into it, the more impressed I am with the abilities of
these seemingly simple animals,” says Randall Beer, an artificial intelligence
researcher at Case Western Reserve University in Cleveland, Ohio. Beer is
collaborating with a team that includes biologists Hillel Chiel and Roy
Ritzman and roboticist Roger Quinn to build robots that control their walking
the way insects do.

Walk tall

“What makes walking interesting is it’s so adaptable,” says Beer. Wheels
are fine if a surface is flat, he points out, but life gets complex for robot
designers when their creations need to creep through holes, climb over
obstacles or scramble up a slope. In these circumstances a walking robot comes
into its own.

But why choose insects as the model for walking robots? “I think it’s a
fairly logical thing when you think about it,” says Beer. “Insects are very
good walkers. In a biped, with two legs, you’re always falling. With six legs,
an insect can adapt its gait so it’s statically stable.”

The key to the new generation of wading robots is a decentralised, “bottom-
up” approach that exploits the latest thinking on insect motor control (“Many
neurons make light work?”, 91av, 7 August 1993). More old-fashioned
robots such as Dante rely on a hierarchical system to control their walking.
Information on leg position and weight distribution has to be fed from sensors
in the leg to a central controller. This controller processes information for
all of the legs, and then instructs each one how to position itself for the
next step.

Spreading the load

The bottom-up approach does away with the central unit and devolves
decision making to the individual legs. Each leg has its own controller, which
moves it according to a number of factors, including the position of the leg
itself and the position and movement of neighbouring legs. Walking patterns
are not imposed from above but come about through a dynamic interaction among
the legs.

This approach comes directly from biologists’ studies of how insects and
other animals walk. Beer credits Keir Pearson who works at the department of
physiology at the University of Alberta in Canada for inspiring his interest
in the area. Pearson carefully observed how insects’ legs move in relation to
one another, and saw that the gait – the legs’ pattern of movement – depended on
the insect’s speed, the terrain it is moving across, what happens when legs
are obstructed or missing, and other factors. Roboticists are now getting
ready to create robots that behave the same way.

When an insect is on the move, one of the factors its legs respond to is
the position of the other legs around it. In mathematical terms, the legs can
be seen as coupled oscillators, whose movements affect one another in the same
way that the motions of two bobbing weights on the same spring interact (see
“The mathematical springs in insect steps’, 91av, 8 October). There
is no central system regulating the legs; instead, an insect’s walking
behaviour is generated by a network of neural connections. This decentralised
approach, which insects evolved over millions of years, is astonishingly
effective. It allows them to adopt various speeds and gaits to deal with
uneven ground and different speeds. They can even walk well after losing one
or two legs.

The movement of an insect’s leg can be broken down into two distinct
strokes. During the power stroke, the leg is in contact with the ground and
propels the insect forwards. This is followed by a return stroke, as the leg
lifts up and swings towards the front of the body. This motion forms the basis
of the insect’s walk, and the part of the insect nervous system that controls
its legs oscillates between two states corresponding to the two strokes.

In a review paper published in January 1990 in Trends in Neurosciences,
Holk Cruse of the University of Bielefeld in Germany took a look at the work
of about a dozen researchers studying cockroaches, stick insects, lobsters and
other invertebrates. Cruse pointed out that when one of a crayfish’s back legs
is performing a power stroke, the leg just ahead of it performs a return
stroke. If something stops the rear leg moving, the leg ahead of it slows down
too.

This simple connection leads to the two typical insect gaits. One is the
relatively slow metachronal wave, in which the rear leg moves, then the middle
leg on the same side, then the front leg, creating a wave-like motion. At the
same time, the legs on the other side of the insect are moving in the same
sequence, but half a step out of phase. The other is the “tripod gait”, which
kicks in when the insect speeds up. In this gait, the front and back leg on
one side are on the ground at the same time as the middle leg on the other
side.

Building on Pearson’s research, Beer designed a computer simulation of a
neural network made of 37 neurons. Unlike a conventional computer, which
depends on explicit, step-by-step instructions to perform its tasks, neural
networks are modelled on the human nervous system, with its multiple
interlinked synapses. In a neural network, interlinked processors respond to
stimuli by “firing” when a stimulus is strong enough, which affects other
processors in the network. Neural networks are especially good where the data
may be “fuzzy”, since they do not depend on clear-cut yes/no decisions, but
make their decisions according to a complex averaging out of all the data they
receive.

Beer’s team did not attempt to reproduce the structure of an insect’s
nervous system, merely to achieve some of the same results. They built a
simple six-legged robot controlled by a neural net, which successfully adopted
metachronal wave and tripod gaits.

For the second robot, Beer and his researchers wanted to build in more
complex behaviours. For the prototype, they switched to a rule-based system to
control the robot as these programs are easier to change than a neural net.
The system is still decentralised, and controls each leg separately according
to the positions of the other legs. But it uses conventional algorithms,
rather than a neural network. Beer says he would prefer to go back to neural
nets in future because they are more robust. For instance, a neural net is
less sensitive to sensor “noise” than an algorithmic controller, and can
operate after some neural connections have been severed.

Lightweight

The body of the rule-based robot was designed and built by Ken Espenschied,
a PhD student at Case Western. The robot is 50 centimetres long, so it fits
comfortably on a table top, and is made mostly of balsa wood and aircraft
plywood to keep its weight down to about 5 kilograms. The six legs are
operated by 6-watt dc motors, three in each leg. The robot’s “brain” is a
personal computer, which is attached to its body by wires.

The designers have given the robot reflexes that real insects use to help
them walk. For instance, the robot automatically keeps its body more or less
level to prevent it tipping over. If a leg hits an obstacle, an “elevator
reflex” kicks in: the leg draws back, lifts up, and steps over the obstacle.
If it doesn’t make contact with the ground when it is expected to, a searching
reflex probes ahead until the leg finds a foothold. Using this reflex, the
robot has managed to walk across a slatted surface in the laboratory.

The next stage will be to build a robot with more sophisticated behaviour
patterns. For instance, Beer would like to build in searching behaviour, or
the ability to follow an edge. Quinn, a roboticist on the Case Western team,
hopes to get the next robot off the table top and out onto the campus to test
it out on real terrain. The team’s main aim is to understand insect walking
and its implications for robots, rather than to build a useful robot.
Nevertheless, the research at Case Western is getting funding from NASA and
the Office of Naval Research, for whom useful robots are the ultimate aim.

One researcher who wants to develop a practical insect-like robot is
Subramanian Venkataraman of the Jet Propulsion Laboratory and the California
Institute of Technology. With funding from the Office of Naval Research, he is
putting together a six-legged robot, and eventually wants to build a flexible
44-legged centipede-like robot. From a practical viewpoint, Venkataraman sees
two advantages to a centipede-like robot. “It can go through tight
clearances,” he says. But also, “if you want to carry bulky, large drums, it
can snake its way around the drum and carry it.”

Venkataraman is using a bottom-up approach to control his robots, though it
is rather different from Beer’s. He points out that the mathematical models
proposed by biologists to produce insect-like gaits also produce many gaits
that insects never adopt. Instead of always switching between metachronal wave
and tripod gaits, they can adopt a weird gait in which some of the legs are
moving twice as fast as others. Venkataraman’s solution is to fashion six
different controllers, each designed to produce a separate gait. As the robot
moves it will automatically choose the most appropriate gait for its speed,
the terrain it is going over, and other factors.

Venkataraman has based many of his ideas on the research of Bob Full, a
neuroscientist at the University of California at Berkeley, who specialises in
insect locomotion. “Cockroaches don’t view these things as a control problem.
They have no notion of stability, of body dynamics,” Venkataraman explains.
Insects use a sense of exoskeletal strain to switch between gaits, he says, so
he has given his robot a flexible spine and the ability to sense the strain
various gaits put on its structure. The idea is that a gait suitable for one
speed and type of terrain may not be so good if the speed or terrain changes.
As the stresses on the robot’s body increase, it will change gaits in order to
relieve those stresses – and in the process move more efficiently. Looking
ahead several years, Venkataraman plans to unify the six mathematical models
into a single model that produces all the desirable gaits and none of the
unwanted ones. He hopes to build this model into a neural network, and use
that to control his robot.

How far can designers go with this approach? Animals exploit their bottom-
up biological designs to perform extremely complex behaviours. However, David
Wettergreen, who designed Dante’s walking controller, predicts that in the
future bottom-up and top-down control will complement one another. Bottom-up
controllers will make robots much better at walking, but he thinks it will be
hard to design a neural network from the bottom up that will give it the
ability to do complex searches, for instance. Those high-level behaviours will
still have to be programmed in centrally.

A helicopter rescued Dante from Mount Spurr’s crater after the robot took
its ignominious tumble. If a successor ever makes a similar descent, it should
be a little more sure-footed, thanks to walking mechanisms its creators learnt
from real live insects.