Page 233 - Introduction to Autonomous Mobile Robots

Chapter 5
218
linear indoor office space filled with real office furniture as obstacles. Traditional sonars
were arranged radially around the robot in a ring. Robots with such sensor configurations
are subject to both tripping over short objects below the ring and to decapitation by tall
objects (such as ledges, shelves, and tables) that are above the ring.
Dervish’s answer to this challenge was to arrange one pair of sonars diagonally upward
to detect ledges and other overhangs. In addition, the diagonal sonar pair also proved to
ably detect tables, enabling the robot to avoid wandering underneath tall tables. The
remaining sonars were clustered in sets of sonars, such that each individual transducer in
the set would be at a slightly varied angle to minimize specularity. Finally, two sonars near
the robot’s base were positioned to detect low obstacles, such as paper cups, on the floor.
We have already noted that the representation provided by the contest organizers was
purely topological, noting the connectivity of hallways and rooms in the office environ-
ment. Thus, it would be appropriate to design Dervish’s perceptual system to detect match-
ing perceptual events: the detection and passage of connections between hallways and
offices.
This abstract perceptual system was implemented by viewing the trajectory of sonar
strikes to the left and right sides of Dervish over time. Interestingly, this perceptual system
would use time alone and no concept of encoder value to trigger perceptual events. Thus,
for instance, when the robot detects a 7 to 17 cm indentation in the width of the hallway for
more than 1 second continuously, a closed door sensory event is triggered. If the sonar
strikes jump well beyond 17 cm for more than 1 second, an open door sensory event trig-
gers.
To reduce coherent reflection sensor noise (see section 4.1.6) associated with Dervish’s
sonars, the robot would track its angle relative to the hallway centerline and completely
suppress sensor events when its angle to the hallway exceeded 9 degrees. Interestingly, this
would result in a conservative perceptual system that frequently misses features, particu-
larly when the hallway is crowded with obstacles that Dervish must negotiate. Once again,
the conservative nature of the perceptual system, and in particular its tendency to issue false
negatives, would point to a probabilistic solution to the localization problem so that a com-
plete trajectory of perceptual inputs could be considered.
Dervish’s environmental representation was a discrete topological map, identical in
abstraction and information to the map provided by the contest organizers. Figure 5.21
depicts a geometric representation of a typical office environment overlaid with the topo-
logical map for the same office environment. Recall that for a topological representation
the key decision involves assignment of nodes and connectivity between nodes (see section
5.5.2). As shown on the left in figure 5.21 Dervish uses a topology in which node bound-
aries are marked primarily by doorways (and hallways and foyers). The topological graph
shown on the right depicts the information captured in the example shown.