PROGRAMMING robots so they can work out what others are thinking could not
only make them much more versatile, but cheaper too.
Manufacturers who buy expensive robots to do specialised jobs like
spray-painting cars would rather buy general-purpose droids that they can train
to do a wide variety of jobs simply by showing them the task. But today鈥檚
copycat robots merely follow a fixed set of coordinates telling them where their
limbs need to move. This means they can鈥檛 handle unexpected events such as
obstacles getting in their way. They go through the motions rather than work out
how to do a job.
Gillian Hayes, George Maistros and Yuval Marom at the University of Edinburgh
say it鈥檚 much better to create robots that can 鈥渦nderstand鈥� the purpose of a
task, rather than just copy mechanical movements. That way, a robot would be
better equipped to cope with interruptions and then carry on when the obstacle,
say, has been removed.
Advertisement
This approach to robotics was just a dream until 1998, when Italian
neuroscientists discovered neural circuits that understand the intention behind
an action (快猫短视频, 27 January 2000, p 22). They found that when
a monkey does something deliberate, like pick a banana off a table, a certain
group of neurons fires in the brain. But the same group of neurons also fired
when the monkey saw another monkey鈥攐r a human鈥攔epeating the action.
Neurons that behave like this have been termed 鈥渕irror neurons鈥�.
Hayes says her group wanted to exploit the way mirror neurons seem to
underpin a type of 鈥渕ind-reading鈥濃�攂y enabling the brain to read the
intention behind someone else鈥檚 movements.
At an AI conference in Massachusetts last week, Maistros and Marom said they
have designed primitive robots with circuits wired in a way they believe is
similar to mirror neurons.
When their robots see a movement that they don鈥檛 recognise, they create a
鈥渘ode鈥� in their processing circuitry鈥攁 point that can either be on or off.
The node represents that movement, and is only switched on either when the robot
sees the corresponding movement, or carries out the movement itself鈥攋ust
like a mirror neuron.
If the point was to lift a glass of water, for example, someone would
approach the table the glass sits on, and this would generate a node in the
robot. So would opening a hand to receive the glass. As the robot sees more and
more of such actions, the nodes they generate cluster into 鈥渢argets鈥� in the
robot鈥檚 brain that represent the overall point of the action鈥攑icking up
the glass. Then to make the robot do the job, you simply activate the target
cluster. It will then use its motors, arms and grippers and so on to perform the
required action.
The full power of the mirror-neuron approach isn鈥檛 obvious in the team鈥檚
prototype robots, because they only learn one action each. But the team now
wants to make the droids more flexible by teaching them that they could move
their gripper forward for a number of reasons, for example when shaking hands,
or to spray-paint an object, say.