A Platform For Vision
Our previous vision guided
robots had crawled about on the floor attached to a computer by an umbilical
which also supplied power and as everybody knows (sooner or later) umbilicals
just get tied in knots, wrapped round wheels, get caught in things, you
think of it - it happens. So Liberator had to be self contained. The plan
was to transmit the video signals to a receiver attached to a computer,
get the computer to figure out what to do and send the control signals
back to Liberator, however that part never got built and so for higher
level processing Liberator still needs to have an umbilical.
Today Liberator consists of
a base unit with two drive wheels and two castors in a wheelchair arrangement.
The two motors came from Autopolisher units (from Halfords?) and each
drive a wheel through a toothed belt reduction drive, power is supplied
from a 12 volt 20 Ah sealed lead acid battery positioned between the drive
wheels. Also on board is the necessary battery charger and a distribution
panel with switches and fuses etc. Around the perimeter of the base are
sensor plates through which Liberator can detect if it is touching something.
Originally the computer was mounted on top of the base unit but it was
moved to above the shoulder plate because I got fed up of crawling on
my knees during software development. All the body is made from WOOD.
Wood is an insulator, you can glue it, screw it, nail it (not advised),
paint it, and saw it and drill holes in it and otherwise shape it without
worrying about the dust shorting out the electric's.
The controlling computer is
a Triangle Digital Services 2020 16 bit ANSI Forth computer which communicates
with the base hardware via an I2C bus, the motors are pwm controlled and
each drive wheel is fitted with an encoder disk for feedback.
At the moment Liberator is
quite happy wandering aimlessly round the lab but really needs someone
who will spend some time on improvements. Please contact The
Shadow Robot Company if you would like to know more or get involved.
The hardest and single most
important aspect of vision robotics is getting a robot to recognise an
object. Most research conducted in this area revolves around creating
methodologies and systems to allow a robot to properly identify objects,
there are of course numerous ways of doing this, but getting them to work
optimally is very hard. In fact some of the most testing examples of Artificial
Intelligence have been developed for intelligent vision systems. No doubt
most cutting edge research into Artificial Intelligence, such as 'Neural
Network Computing,' will still remain within the realm of Vision Systems
for some time.
However progress has been made
and one methodology, 'Segmentation,' although simpler than some advanced
Artificial Intelligence Systems, proved to be an effective starting point
for engineers at Shadow. Peter Holman, a Shadow Member, has worked extensively
on vision systems and set about working on a Segmentation system for Liberator.
Segmentation as a method means breaking down the picture of an object
into discrete parts. These parts are effectively parameters or segments
by which the most significant of these can be used to identify an object.
Therefore an object is defined by the Robot in recognising certain parameters.
However the Robot must gain knowledge of object parameters in order to
compare the object it is seeing with the one stored in its database in
order to make a positive identification. The Robot is fed information
about numerous parameters - segments - of an object which it then stores
in its database. The Robot becomes 'aware' of an object when it is presented
it in ideal conditions. One of these 'Significant Segments,' could be
an object's hue or surface area.
© David Buckley, Peter
Holman & Rufus Wood 1999