快猫短视频

CHECKING sewers for leaks, cracks and blockages is a task most of us wouldn鈥檛
touch with a bargepole. But soon we won鈥檛 need to, as a team at the University
of Manchester has come up with a way to teach robots how to do it.

An engineer will normally inspect a sewer, or alternatively, send a
remote-controlled video camera trundling through it and watch the pictures above
ground. Either way, it鈥檚 time consuming and unpleasant. But Stephen Marsland,
Ulrich Nehmzow and Jonathan Shapiro at Manchester鈥檚 department of computer
science say that a robot designed to spot anything out of the ordinary could do
the job just as well.

鈥淥ur goal was to build a robot that spots things it doesn鈥檛 know,鈥� says
Marsland. Their work was inspired by the phenomenon of habituation, where
organisms become desensitised to familiar stimuli.

The team designed a neural network that ignores familiar stimuli and homes in
on new ones, such as lights flashing at different rates or from different
directions. They flashed lights at a robot fitted with a number of light
sensors, and found that the robot learned to ignore constantly occurring stimuli
but would turn to face new stimuli鈥攆or example, when the direction or
frequency of the flashing changed.

Now that they have a robot that can recognise novelty, the group is working
on an algorithm that works with images. Marsland says the idea could be
particularly useful for inspection tasks. 鈥淵ou鈥檇 train the robot by sending it
through a sewer that you knew had no cracks or blockages. But when it saw
anything it did not recognise鈥攕uch as tree roots or cracks鈥攊t would
immediately issue a warning,鈥� he says.