Page 24 - Designing Autonomous Mobile Robots : Inside the Mindo f an Intellegent Machine

Measure Twice, Cut Once

So what could go wrong? The answer is just about everything! The first problem has
to do with timing. If the task that gets the sonar ranges is actually running the sonar,
or it must wait for the ranges to be determined by another system, then the length of
time required to process the range information will vary depending on conditions.
Furthermore, the calculations of the various states will take variable amounts of
time. The reason that this is important is that the states must calculate accelerations
and velocities, and to do this they must be related to a consistent time base.
The final task of running the motors will be executing a control program that will
also require consistent timing. It is therefore necessary that the program loop at a
constant and known rate so that these calculations can be made. Reading a hard-
ware clock is not an option, as the sampling rate of the system needs to be constant
to assure stability. If the reader is familiar with filter theory, all of this will sound
very familiar! The simplest solution to the problem of achieving constant timing is
to eliminate the loop back to the top of the process, and execute the whole process
from a timer interrupt.
If we assume that the process is timer driven, it will be necessary to assure that the
task is completed before the timer fires again. If the timer starts again before calcula-
tions are complete, then partially calculated parameters may be overwritten and
contaminated. A task like this one is thus said to be nonreentrant. Worse yet, re-
peatedly reentering a nonreentrant task can cause the processor to eventually
overflow its stack and crash. This can cause very strange and unwelcome behavior in
our little appliance.

Next consider the “straw man” scenario of Figure 1.3. Here our robot encounters a
target on its left side (State 3), begins turning right, drives into the path of a target
on that side at an even closer range (State 4). As it turns away from the target on
the right, it goes back into State 3. The result is that the robot drives forward so that
the closer target approaches along the right edge of the right transducer pattern
until the robot comes to a halt, or worse, sideswipes the closer target (Figure 1.4).
It has already become apparent that our control architecture is flawed. We must
have a strategy for escaping from such a situation. One problem with our strategy is
that we are trying to solve the problem instantaneously, without regard to the recent
past. As soon as we begin to turn away from the closer target, we forget all about it,
and react to the further target again. Truly, a robot that doesn’t learn from history is
doomed to repeat it!

7