Page 237 - Designing Autonomous Mobile Robots : Inside the Mindo f an Intellegent Machine

Chapter 15

The first problem was one of position accuracy. We quickly learned that recording
“bread crumb” paths was a bad idea for indoor robots, because any unintentional
weaving would be recorded and duplicated by the robot. The size of the data re-
quired for bread crumb paths was also orders of magnitude larger than for other
methods.

True, the data generated by walk-behind teaching could be massaged by line-fitting
programs to generate smoother curves and straighter lines, but there is still a prob-
lem. Anyone who has used a drawing program that performs curve fitting (as lines
are drawn using a mouse) can quickly identify with the problem of this approach.
The algorithm often makes wrong assumptions about the artist’s intentions.
For example, under most circumstances a robot running along a wall will maintain a
constant distance from the wall, so it is logical for the learning process to attempt to
massage the bread crumb data to accomplish this. However, there are times when
the path must violate this assumption. To communicate these subtleties to the
learning program, the walk-behind teacher is essentially trying to accomplish the
same thing as the aforementioned artist, but with a 500-pound mouse!

Distance issues

Another issue that is often not appreciated is the sheer size of some environments.
Many of the installations we were to program were more than a million square feet
in size with tens of miles of paths. In most applications, there is no return on invest-
ment by putting a robot into a small area, so all but the most modest installations
represented a marathon effort to program.
The choice is between walking twenty miles while arguing with a software interface
about what you really wanted to teach it, or sitting at home and drawing paths on a
map. It should be obvious that the walk-behind approach is not much fun. Road-fol-
lowing vehicles that can be ridden while being taught are another matter however.

Data management issues

In Chapter 7, we discussed the concept of building maps from scratch vs. starting with a
high quality drawing. A good facility drawing has information about much more
than just walls and doors, and shows things that no mapping program can determine.
Any good automation engineer knows the golden rule:

Never throw away useful information.

220