Page 431 - Introduction to AI Robotics

414
11
Localization and Map Making
depends on the density of air, so if a robot is going from sea level to high
mountains, an adjustment factor can be added to the raw readings. Another
approach is to change the threshold on what the robot considers a “really
occupied” region. Lowering the threshold makes the interpretation of the oc-
cupancy map more conservative; more occupied regions that may be phan-
toms are treated as if they were real. This typically doesn’t work well in
cluttered or narrow environments because the robot can get blocked in by
false readings. Increasing the threshold can make the robot less sensitive to
small occupied regions which may not get many readings. Finally, a com-
mon solution is to slow the robot’s velocity down; however, this exacerbates
sensor noise in HIMM/GRO updating mechanisms and to a lesser degree in
Bayesian and Dempster-Shafer.
Other possibilities for tuning the performance include changing the sonar
TUNING BAYESIAN model and the update rules. In practice, only two aspects of the Bayesian
MODEL sonar model are tuned: the field of view and the prior probability that an
area is occupied. In difficult environments, the range R accepted as valid is
often shortened. A robot might treat a range reading greater than 4 feet as
being empty even though the sonar range is theoretically covers 25 feet or
more. The rationale is that the likelihood that long readings are accurate is
small and the robot is more interested in obstacles nearby. Of course, this
can limit the robot’s maximum safe velocity since it may be able to cover
a distance faster than it can determine reliably that there is anything in it.
Likewise, the for the field of view is often adjusted. In Sec. 11.3, the prior
probability was assumed to be P (H) = P (:H) = 0:5. However, this isn’t
necessarily true. In some cases, the area to be covered is actually more likely
to be occupied. Consider a robot operating in a narrow hallway. Compare
the hallway to the area that can be covered by the robots sonars. Most of
the field of view is likely to be occupied, which may argue for a P (H)
P (:H). Moravec’s ongoing work in sonar-based occupancy grids has shown
improvement based on using more accurate priors. However, this requires
the robot or designer to gather data in advance of the robot being able to use
the data. There is work in adaptive learning of the parameters.
TUNING DS MODEL Dempster-Shafer theoretic methods have less to tune. Priors are not re-
quired as with Bayesian; Dempster-Shafer assigns all unsensed space a belief
= information, the appropriate expecta-
of m(dontknow ) :0. If there is prior 1
tions can be placed into the grid. However, this is rarely if ever done. Tuning
with Dempster-Shafer consists primarily of changing the field of view pa-
rameters, and R.
TUNING HIMM HIMM/GRO have many more parameters that can be tuned, which can