TEACHING a robot to stay upright is as easy as learning to ride a bicycle or
a unicycle, according to an international team of robotics researchers. The
researchers have taught a robotic wheel to recover its balance by imitating the
way that people learn to ride a bicycle.
Called Gyrover, the wheel was developed by researchers at Carnegie Mellon
University in Pittsburgh, Pennsylvania. It consists of little more than a small,
bulbous wheel driven by an internal motor and it steers itself by tilting a
gyroscope.
One advantage of this design is that it is very stable at speed, just like a
bicycle, allowing it to stay upright while travelling fast over rough terrain.
The robot鈥檚 stability could make it ideal for exploring other planets, say its
creators.
Advertisement
The bulbous wheel design means that the Gyrover can never flip upside down,
unlike other robots, says its developer Ben Brown. He says the design was
inspired by childhood memories of rolling tyres downhill.
But like a rolling tyre, the Gyrover can still fall over when it loses its
stability, says Brown, so it needs a way to right itself. To do this
automatically would normally require a highly complex computer model. Such a
model would be extremely difficult to design, says Yang Sheng Xu at the Chinese
University of Hong Kong, who developed the Gyrover鈥檚 control system.
So Xu decided to harness the balancing skills that people acquire when they
learn to ride bicycles. Why design it when you can copy it, he says. 鈥淧eople
have shown themselves to be extremely adept at mastering the complex and
difficult control of dynamically stable systems,鈥� says Xu. 鈥淎 skilled operator
can control the tilt-up motion of the robot very well.鈥�
Together with Samuel Au and Wilson Yu, he designed a neural network鈥攁
computer program that can imitate the way the human brain learns鈥攖o
control the robots. To teach the neural network what to do, a human operator
used two joysticks: one to control the motor and the other to steer the robot by
tilting the angle of the gyroscope.
While the operator controlled the Gyrover, sensors on the robot monitored its
stability. The neural network then married up the inputs from the joysticks with
information from the robot鈥檚 onboard sensors, to examine how the operator
responded when the robot was in danger of losing its balance and learn to
imitate these reactions. From this information it created a computer model that
enables it to control the robot unaided.