IEEE Robotics & Automation Magazine - September 2017 - 60

2
4
6
Number of Fires

Extinguished Fires

Average Patrol Rounds

Extinguished Fires (%)

100

Average Patrol Rounds

Figure 8. The test-bed simulations for single-robot surveillance
with increasing random fire events at the coffee room. Average
values for the percentage of extinguished fires and the number
of patrol rounds of the robot are shown.

only stopped them to attend and extinguish fires. As expected, as there were more fires, the robot was able to perform
fewer rounds. Besides, some fires had been triggered close to
the end of the experiment, leaving the robot with no time to
reach the coffee room. Therefore, as the number of fires
increased, the extinguishing rate gradually degraded.
Deployment in a Shopping Center
We performed a similar experiment in a more realistic environment located in a shopping mall. As a first step toward
integration, we deployed our system in the technical corridors
beneath the mall that are closed to the public. The map of the
scenario and its topological abstraction are shown in Figure 9.
Here, in addition to waving events, trespassing events were
introduced. (A video summarizing the tests performed can be
viewed at https://youtu.be/Ivx908SSzlk.)
In this scenario, three functionalities of the system were
tested to assess its capabilities to respond to different situa-

East Transit Zone

East Park

tions using a single balanced MDP policy. In the absence of
events, the robot began moving around the environment,
selecting the next area to visit among those defined in Figure 9(a) and ensuring that key areas were visited frequently.
During the robot's patrol, we triggered random trespassing
events by entering the restricted technical corridor (see Figure 6). Each time, the robot stopped its patrol, its policy dictating that it move toward the intruder's detected position to
intervene. Upon arrival, the robot requested the intruder to
"leave the area immediately." After he was gone, the robot
resumed its patrol. We also triggered waving events to test the
robot's ability to perform visitor assistance. These tests consisted of a person entering a camera's field of view and hand-waving to request help. In response to the waving detection, the
robot stopped patrolling and went to the position of the event
to interact with the visitor, prompting the person to select
among several possible areas in the environment. Once the
visitor selected a desired destination, the robot led the way.
We carried out a third deployment of our multiagent surveillance system in the commercial, publicly accessible areas
of the same shopping mall (see Figure 10). The functionalities
and behaviors obtained were qualitatively identical, but the
autonomous navigation of the robot was made considerably
more difficult due to the characteristics of the environment
and the robot's hardware limitations (e.g., the glass panes of
storefronts sometimes eluded its laser range finder).
Event-Driven POMDPs for Multirobot Surveillance
In the next experiments, we adopted an alternative decisionmaking approach suitable for multirobot settings with partial
observability of event occurrences. In contrast to the MDP
model, a POMDP explicitly considered that the event detector (and hence robots' observations) was susceptible to errors.
Such errors may come in the form of false positive detections
(e.g., incorrectly detecting a person in an empty room) or
false negative detections (e.g., failing to detect a person).

East Park

East Transit Zone

East Door
Central
Long Transit
Corridor Zone

Long Corridor

East Door

Central Transit Zone
Stairwell West Door
Foyer

West
Transit
Zone

West Park

Foyer

West Door
West Transit Zone

Stairwell

West Park
(a)

(b)

Figure 9. The topological map used at the shopping mall. (a) The defined selected areas designated for the robot to visit. (b) The
topological connections.

IEEE ROBOTICS & AUTOMATION MAGAZINE

SEPTEMBER 2017

https://www.youtu.be/Ivx908SSzlk

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2017