DPRG mailing list archives: [DPRG] Are representations what makes Can Can hard?

Tue Dec 5 17:56:11 CST 2006

On 12/5/06, Randy M. Dumse <rmd at newmicros.com> wrote:
> Why is Can Can so hard? Is it fundamentally a different kind of
> contests from the others?

Warning: some of the following arguments apply only to the "full-size"
Can-Can rules and are not applicable to the new "travel-size" version.

In the other contests, a robot is required to consider either the
floor color or the "occupiedness" of space and then act on that
information is interesting ways.  Can-Can's primary decision point
hinges on the presence or absence of a can, and requires robots to do
interesting things with that information.  On their surface, these
seem to be very similar.

What makes Can-Can so difficult is that it moves beyond the realm of
sensing and into semantics.  There is not, and will probably never be,
a "can sensor" that will produce a signal when a soda can is present.
Instead, a robot must deduce, infer, or guess whether the sensor
readings represent a can, which should be retrieved, or an obstacle,
which should be avoided.

Obstacles (walls) and cans all look very similar to the sensors that a
robot has.  In our contest, they are both:
  - white,
  - vertical, and
  - approximately the same height

The only practical things that our robots can use to determine whether
something is a wall or a can (that I can identify) are:
  - curvature (watch out for inside corners),
  - position,
  - size (really the same thing as curvature),
  - navigability (Can the robot travel all the way around it
        (really the same thing as position)), and
  - mass.

These are all properties of extended objects, so it is unlikely that
we will have sensors to measure these things directly.  Thus, any
robot that wishes to distinguish a can from a wall needs to have some
fairly complex method to do so.  David managed to find a resonable way
to do this with SR04; it uses a pair of distance sensors to determine
the size (or is it the curvature?) of an object, and tries to grab
anything that looks like a can.  If the object then doesn't behave
like a can (doesn't allow the gripper to close), SR04 decides that it
isn't a can after all.

Also, if you don't know where a can is, you have to do some kind of
search pattern to find one.  In your search pattern, you need to be
careful to exclude any cans you've already found, which look even more
like target cans than walls do.

Compare that to the simplest known way to do quick trip: Go forward
until you hit something, and then go backward until you hit something.

In conclusion, Can-Can fundamentally requires a MUCH more
sophisticated behavior than any of the other contests, especially if
you don't want to put Can-Can-specific sensors on your robot.

  -- Eric Sumner