Page 430 - Introduction to AI Robotics

413
11.6 Comparison of Methods
or crosstalk, obstacles might be missed or appear at the wrong distance. If
the robot is stationary, HIMM will generate high belief in an incorrect map.
Bayesian and Dempster-Shafer theory also suffer from the same defect. Since
they usually cover a larger area, the problem with gaps in walls is usually
avoided. But problems with phantom readings still cause incorrect maps.
The plots of the rate of accrual of belief show that multiple identical read-
ings will cause the robot to quickly believe that its occupancy grid is correct.
Once P (HjS) or m(H) reach 1.0, there is no revision downward. HIMM can
revise belief because it subtracts strictly based on the current reading. But
HIMM must have a new, contradictory reading to cause this to happen.
The reason Bayesian and Dempster-Shafer methods degenerate when the
robot is stationary and receives multiple, identical readings is because the
assumption that the observations are independent has been violated. If the
robot is at the same location sensing the same object, the value of reading
. Since the robot hasn’t moved, the ob-
S t n+1 is likely to be the same as S t n
servations cannot be considered to be taken from two different experiments
or by two different observers. This serves as a cautionary note about making
simplifying assumptions: it is important to understand when those assump-
tions lead to counterproductive results.

11.6.4 Tuning
Fig. 11.14 shows the performance of the three updating methods for a hall-
way with significant specular reflection. All three methods show the hallway
as being wider than it really is. This would be a serious problem for naviga-
tion and obstacle avoidance. The sensor noise was not eliminated by the use
of an occupancy grid. In many cases, a large amount of sensor noise can be
eliminated by tuning the model and updating algorithms.
Therefore an important criterion for an algorithm is how easily it can be
tuned for a particular environment. For example, in environments which

provoke a high degree of specular reflection in sonars, a < 8 is often
used to reduce the registration of noise in the occupancy grid. Why put false
readings into the grid over a large area that will take several contradictory
readings to eliminate? It can often take one or more days to tune a set of
sonars which were producing near perfect occupancy grids in a laboratory
for a new building.
Occupancy grids can be tuned for a task environment in at least three
3 WAYS TO TUNE ways. One way is to leave all the algorithms the same but concentrate on
adjusting the physical equipment. For example, the time of flight of sound