The Multi-Robot Patrolling Problem

Patrolling an infrastructure with multiple robots is no different than other multi- robot assignments, in the sense that it incorporates all the previously mentioned characteristics of MRS. To understand this problem, it is important to firstly introduce the definition of patrol.

Definition 1 (Patrol). According to the Webster’s online dictionary [Webster’s, 2013], to patrol is literally “the activity of going around or through an area at regular intervals for security purposes”.

In the context of this thesis, it is assumed that instead of having a human guard or a group of men, the patrolling task should be performed by multiple autonomous and cooperating mobile robots in a real-life environment.

1.3. The Multi-Robot Patrolling Problem 7

Patrolling Problem

Surveillance Enemy Detection

Perimeter Patrol Area Patrol Pursuit- Evasion Adversarial Patrol

Figure 1.2: Research areas inside the Patrolling problem.

Patrolling is as a somehow complex multi-robot mission, requiring an arbitrary number of agents to coordinate their decision-making with the ultimate goal of achieving optimal group performance. It also aims at monitoring, protecting and supervising environments, obtaining information, searching for objects and detect- ing anomalies in order to guard the grounds from intrusion. Hence, a wide range of applications are possible, as exemplified in Table1.1.

It is the author’s belief that employing teams of robots for active surveillance tasks has several advantages over, for instance, a camera-based passive surveillance system. Robots are mobile and have the ability to travel in the field, collect environmental samples, act or trigger remote alarm systems and inspect places that can be hard for static cameras to capture. These capabilities are greatly beneficial to safeguard human lives and in terms of the flexibility of the deployed system..

Essentially, investigation on the Multi-Robot Patrolling Problem (MRPP) is divided in two classes, as seen in Figure 1.2, surveillance and enemy detection. Both have different goals and metrics. While a good strategy for detecting enemies does not necessarily require constant visits to all locations of the environment, a good surveillance strategy (also known as supervision or monitoring) intrinsically

involves such frequent visits to all places in the environment1.

On one hand, enemy detection techniques are usually divided in two types of strategies in the literature: adversarial patrol [Basilico et al., 2009, Agmon et al., 2011, Aguirre and Taboada, 2012] and pursuit-evasion [Ishiwaka et al., 2003, Vieira et al., 2009, Strom et al., 2010]. Both of these assume a target or an intruder inside the environment. However, while the target is mobile in the latter one, in the former one the target may be motionless and an explicit search for the opponent may not exist. For example, agents in an adversarial patrol task may strategically position themselves in the environment to attain maximal environmental coverage. Thus, the problem finds inspiration from the classical art gallery problem [Sherman, 1992], which is a visibility problem that aims to find the minimum number of static guards who together can observe an entire polygonal region. The goal in enemy detection is to recognize the target as quickly as possible. Usually behavior models of the adversary are considered, and performance is evaluated according to exploration time and other similar metrics.

On the other hand, surveillance techniques can be divided in perimeter patrol and area patrol. As their name suggests, perimeter patrol addresses supervision around a closed area [Agmon et al., 2008, Marino et al., 2009, Jensen et al., 2011], while area patrol is concerned with supervision inside a region of space [Machado et al., 2003, Iocchi et al., 2011, Pasqualetti et al., 2012a]. Oppositely to enemy detection, in surveillance methods, agents are continuously traveling inside an area for an indefinite period of time, seeing that every position in the environment, or at least the ones that require surveillance, must be regularly visited. Therefore, metrics based on the time spent without visiting important places in the environment or the frequency of visits have been proposed to gauge the performance of several strategies. Note that an intruder may not necessarily exist in these scenarios, e.g., cooperative cleaning.

The focus in this thesis is on monitoring and supervision of environments, more specifically on area patrolling missions. Therefore, the words “Patrol” and “Patrolling”, are implicitly used in this sense throughout this document.

1_{Alternatively, it may involve visiting all important designated locations instead of visiting} every single location.

1.3. The Multi-Robot Patrolling Problem 9

Some distinct area patrolling strategies for teams of multiple robots have been presented in the last decade and it is consensual that a good strategy should minimize the time lag between visits in strategic places of the environment. More detail on the literature is given in the next chapter.

Also, it is important to address the characteristics of the patrolled environment. For example, the environment may be static or dynamic, i.e., there may be changes in the environment during execution of the mission due, for example, to mobile obstacles. In these cases, the agents have to keep track of all changes and update their local representation of the environment. There may be also areas of the environment with different patrolling priorities if, for example, a region is more critical or more susceptible to attacks than others. In this case, such regions will need to be visited more often.

Robotic agents are normally endowed with a representation of the environment, which is typically an occupancy grid model, which in turn, is normally abstracted by a simpler, yet precise representation: a topological map (i.e., a graph).

As it is shown in the next chapter, most of the previous works in this field are based on topological representations of the environment. By having a graph representation, one can use vertices to represent specific locations and edges to represent the connectivity between these locations. The multi-robot patrolling problem can, thus, be reduced to coordinate robots in order to visit all vertices of the graph ensuring the absence of intruders or other abnormal situations, with respect to a predefined optimization criterion.

Beyond the representation of the environment, it is important to consider other properties of the robotic team like their perception, which is not necessarily global. Agents may only have local awareness of the environment around them, which, in general, makes the problem harder to tackle and more dependent on communica- tion mechanisms, in order to update knowledge about the state of the environment and teammates.

Moreover, it is important to define such communication mechanism. Agents may need to share state information, communicate their intentions, negotiate pa-

trolling regions with other agents or exchange other information that might be important, considering the strategy proposed to better achieve their goals or the team’s goals. With this aim, they must respect a communication protocol, which can be accomplished, for example, through explicit peer-to-peer messages between them or using a blackboard scheme. Nevertheless, some strategies proposed in the literature do not use inter-robot communication at all. For example, a central- ized coordinator unit may compute a set of patrolling routes a priori and simply assign each local trajectory to each robot so as to cover the whole environment. Regarding coordination, these strategies are called centralized. Those that do not rely on a central unit are called decentralized or distributed.

In distributed strategies, agents may have a reactive behavior, simply inter- acting with the immediate surroundings or may have enhanced capabilities, like autonomous decision-making. In such architectures, agents should continuously decide where to move next after clearing each location. As a consequence, dis- tributed strategies generally benefit from greater robustness to agents’ failures, due to distribution of intelligence among the components of the system, as shown in this thesis later on.

In addition, on any given team of robots, agents capabilities is an important issue. For example, in a centralized strategy, it may be appropriate to have a heterogeneous team of robots, where the central coordinator would assign different specific tasks according to each agent’s distinct capabilities. Using a homogeneous or heterogeneous architecture is a decision that relies on the actual cost of the multi-robot system, the application domain and the intended performance of the team. It is yet left to be proven which organization is advantageous in this context.

In this work, it is foremost studied distributed patrolling architectures with robots endowed with local perception capabilities, in environments with fixed topology (though not necessarily static), their design, effectiveness, potential to scale to larger team sizes, as well as their application in realistic scenarios. Sec- ondary questions, like dealing with intruders or monitoring topological changes in the environment were not addressed in this thesis and are left as future work.