91av

IMAGINE the scene. A roving robot is trundling over the Martian landscape
when static electricity zaps one of its sensors, blinding it. It’s millions of
miles from the repair shop back home, so what can it do to recover?
Simple—it rewires its nervous system so other sensors can take over from
the dead one, then carries on.

That’s the idea behind a robot controller being developed at the department
of electronics and computer science at the University of Southampton. It is
inspired by the ability of animals to adapt to injury and unfamiliar
environments by rearranging the connections in their brains.

The brain’s ability to change the connections between its neurons is known as
“plasticity”, and is crucial to the way animals learn how to respond to sensory
inputs. The Southampton team has built plasticity into the software that
controls a pair of robots.

Terry Elliot and Nigel Shadbolt built their biologically inspired plasticity
algorithm into a simple two-wheeled lab robot. The droid was equipped with eight
infrared sensors, and its “nervous system” consisted of a neural-network
program. The program learns from experience how to interpret signals from the
sensors, and controls the motors driving the two wheels so it can avoid bumping
into obstacles.

The algorithm is based on a biological model in which “growth factor”
chemicals strengthen the influence of the most active synapses—the
connections between neurons. The growth factor is in short supply, so neurons
are forced to compete for it. “If a neuron takes up that growth factor then that
means there’s less left for other synapses,” explains Elliot. But if a neuron
stops taking up growth factors, that leaves more for its near neighbours.

In the robot, the growth factor is represented by a fixed numerical value
that has to be shared between the connections to its sensors. When one sensor is
damaged, its neighbours get a bigger share—strengthening their influence
over the robot’s movement. “The robot is recovering performance because it is
allowing its nervous system to be plastic when part of its sensors have been
knocked out,” say the researchers.

To see how well the robot adapts to damage, the team turned off some of their
sensors. “Without the plasticity algorithm the robots obviously crash more
because some of its sensors have gone,” says Elliot. But when running the
algorithm, the number of crashes halved as the robots rewired their brains to
make the best possible use of the remaining sensors.

This could be very important for robots placed in unpredictable environments,
says Phil Husbands at the University of Sussex. For example, on distant planets
it might be useful to navigate visually during the day and with infrared sensors
at night. “It is very hard to do with conventional algorithms,” says Husbands.
“You soon get in a mess.”

While there are other means of getting software to repair itself, they are
nowhere near as good as the ways nature deals with the unpredictable. “Biology
has had about 700 million years for evolving organisms that have had to adapt to
their environments,” says Elliot. “So it has some useful tricks for doing that.”