[From Bill Powers (2005.04.25.MDT)]
Brian D’Agostino (2005.04.24.2225) –
I need
to be able to build robots using Lego parts and a Handy Board processor,
with code written in Interactive C. The robots need to be able to
home in on and pick up grey plastic rings and put them in one bin, and to
home in on and pick up green plastic bottles and put them into a
different bin. I have analog light and IR reflectance sensors, an
optical range finder, and digital touch and slotted encoder sensors, as
well as servos and motors. Would appreciate any ideas and
references any of you may have . .
This problem has several parts, and it would be good to tackle
them one at a time. You need a way to alter the (1) speed of motion and
the (2) direction of motion. You need a way to (3) pick up and let go of
an object. You need a way to (4) sense the direction of an object
relative to the arm that picks it up.
I assume that the Lego kits provide solutions to (1) and (2), and
possibly (3). So if I were doing this project, I would set several
groups to work on providing interfaces between the program and the
actions. This is like designing the lower-order systems that will be used
by the higher order systems. For example, the higher order system will
need to be able to adjust a number that determines speed (rate of change
of position), and another number that adjusts turn rate (rate of change
of direction). Speed is controlled by changing the voltage that drives a
motor; rate of change of direction is controlled by changing the angle of
a steerable wheel. If you get that workling, the next level of system in
the program will be able to control its perceptions just by changing the
relevant numbers. If you have sensors for speed and direction you can
make the lower-order systems into control systems; otherwise just
concentrate on making them reliable and smooth open-loop
actuators.
The hardest problem is (4), sensing the objects. Actually this has
several parts. First you have to figure out how to discriminate the
objects (including the bins) from the background. What are the
characteristics of the objects that you have sensors to sense? You
mention color and shape. But there is also position (direction and
distance) relative to the pickup device. If you can devise a way to
create a perceptual signal that indicates the left-right position of the
object relative to the left-right position of the pickup device and the
position of the pickup device relative to the position of a bin, you will
have most of the problem solved. All that will remain will be to
alter speed and direction until the sensed distance between the pickup
device and an object is zero, actuate the picker-upper, alter speed and
direction until the sensed distance between the pickup device and a bin
is zero, and turn off the picker-upper.
So how can you discriminate the object from its background? Sensing shape
is probably out of the question. You’d need a video camera to create a
visual map of the scene, and then a very sophisticated program which I
don’t believe anyone knows how to write to pick out a ring shape or a
bottle shape (in any orientation and at any distance) from the
background. I assume that what you have are photocells which just detect
light intensity.
First you have to make the sensing directional, so the photocell sees
light only from one direction. You can do that by using a lens, or less
efficiently by letting it look through a tube blackened on the inside.
Then you need a way to tell left from right, which means using two
photocells receiving light from slightly different horizontal directions.
If the signals from the photocells are representable by numbers (not just
on-off), you can find the exact direction to a bright light by using the
left-right movements of the robot to make the two signals equal (make the
difference between signal magnitudes zero).
If you detect the object by its color, you will have the photocells
looking through filters that allow only that color through. Then you
wheel the robot around in a circle two or three times, measuring the
intensities of light from the two photocells and picking out the maximum
intensities. That allows you to set a reference level, after which you
wheel around until the sensed intensities match it. The robot is then
pointing toward the object , or at least to the place where that color is
the most intense. Moving forward will then bring the object into contact
with the picker-upper if you have placed everything correctly (move
forward until the contact sensor indicates non-zero).
If you’re very clever you can get a crude kind of binocular vision out of
this – suppose you could swivel one of the sighting tubes a little bit,
say up to 20 degrees, relative to the other, and measure the angle when
the program thinks they are seeing the same thing.
I don’t know enough of the details of your equipment and setup to be more
helpful. But the obvious main problem is how to pick out an object from
its background and sense its direction relative to the heading of
the robot. That gives you a perceptual signal to control, and controlling
it will then be relatively easy.
Best,
Bill P.
···