May 2, 2006— Using a program that mimics the way human neurons control reflexes, a
robot has achieved the fastest gait yet of a two-legged machine.
The biologically inspired computer models not only give the so-called
RunBot a stable gait, but also reinforce its learning to achieve faster
speeds.
Developed by Tao Geng at the University of Stirling in Scotland, the
9-inch (23-centimeter) tall machine can't run yet, but it is able to speed up
from a slow lumber to a fast clip in less than three minutes and
achieve a gait comparable to the fastest relative speed of a person
walking.
Unlike most other bipedal robots, RunBot doesn't provide a computer with intensive feedback on its stability and motion.
Instead, RunBot has just a handful of sensors in its feet and
hips, as well as software-driven motor neurons, which initiate reflexes according to
information gathered by the sensors.
It does not have movement in its ankles, but the bottoms of its feet
are curved, with only one point touching the ground.
As soon as a foot touches the ground, a neuron initiates a reflex in
the hips and the knee joint of the opposite leg, bending it and
swinging it forward.
A sensor in the hip measures the angle of the
forward-swinging leg and causes it to straighten out in time to hit
the ground.
"You coordinate these reflexes so they produce oscillations. the one
leg touches the ground, the robot falls forward onto the other leg.
It's like clockwork," said Florentin Wörgötter of the
University of Göttingen, Germany.
Wörgötter, along with professor Bernd
Porr of the University of Glasgow, contribute to the research and serve
as Geng's advisors.
Software that mimics neuronal control allows RunBot to optimize two
parameters: knee bending and hip swinging. When the machine begins
walking, it starts out slowly, more obviously bending its knees, which
is what humans do, said Wörgötter.
The neuronal control allows the robot to try different things. So as
it goes faster, it may do something like avoid bending its knees to
swing the leg forward faster. By learning how to adjust the stride to
the pace, RunBot keeps its balance.
"The robot's learning does not involve falling, which is a frequent
problem in bipeds. One could imagine other quantities being improved,
such as energy efficiency, with a similar approach," said
Jessy Grizzle of the University of Michigan.
Grizzle is part of team
that developed a bipedal robot that could run a couple of steps.
The downside of Geng's approach, said Grizzle, is that its control
algorithm is not guaranteed to work right off the bat.
"To understand what I am getting at, imagine how you would feel about
a car that you had to drive around for a few weeks before it ran
well," said Grizzle.
Wörgötter said that walking isn't so much the issue as the robot's
ability to learn and adapt.
He and his team are looking ahead to
adding sensors, such as range finders, to make the robot more aware of
obstacles and to make RunBot actually run.
