**3. Firefighting Robots**

Once we have analyzed the current state of firefighting and the opinions of professionals, we must address a new question: "can technology help to solve any of the presented problems?" This section collects the most relevant works that apply robotic and automation technologies to firefighting activities. Our analysis focuses on multi-robot systems and aerial robots used for the prevention, surveillance, and extinguishing of forest fires. However, relevant works that propose other types of robots and consider urban or indoor scenarios are also featured.

As previously occurred with industry, agriculture, and services, robots are being applied to intervene in emergencies and, more specifically, to fight against fires. According to our survey, firefighters are receptive to these technologies when they support their work and do not change its conditions. A previous study with fire chiefs of New Jersey (United States of America) supports these conclusions: they are willing to use drones in firefighting operations, but they point out budget, manpower, and regulation issues [8]. The public opinion about the use of drones for cargo, passenger, and commercial transportation is analyzed by [9], including explicitly firefighting in this last group of applications. The participants of this study support the use of drones for cargo and commercial applications, but they prefer piloted aircrafts for passenger transportation. Finally, a comprehensive survey on the public opinion about drones considering multiple applications and risks can be found in [10].

As most of the relevant articles focus on one or a few specific tasks, we have classified them into prevention (Section 3.1), surveillance (Section 3.2), and extinguishing

(Section 3.3). This classification is supported by several papers in the literature: for instance, Ref. [11] distinguishes between activities before fire (vegetation mapping, surveillance, and risk estimation), during fire (detection and extinguishing), and after fire (ember search and damage assessment).

#### *3.1. Prevention*

As already explained, prevention is considered the first step of firefighting and encompasses two classes of actions. The first ones involve social activities that seek to prevent fires from occurring, usually developing awareness campaigns targeting key groups as farmers or tourists. The second ones group multiple works on vegetation to reduce the risk of fire and generate discontinuities to difficult their propagation. Logically, the potential of robots to improve current results is higher in these latter operations. The preparation of vegetation is an activity that requires remarkable efforts, where a lack of human and material resources is perceived. Robotic technologies can make this activity more efficient in two ways.

On the one hand, drones can take aerial images that can be used to plan these tasks: detecting the most problematic areas, selecting the vegetation to remove, planning routes for its extraction, etc. Some techniques developed for precision agriculture can be applied in this context [12], such as the detection and identification of plants and trees in highresolution images [13], three-dimensional LIDAR scans [14], and multispectral images [15] acquired by drones. In all these tasks, drones have been revealed as a suitable alternative to satellites, since they offer greater availability at a lower cost, as well as they are less dependent on the weather conditions in the area of interest [16].

On the other hand, ground robots can support the activities aimed at remove vegetation in forests, playing an intermediate role between the manual labor of firefighters and the heavy machinery used by them. These robots can reach a compromise between the flexibility and precision of firefighters and the quickness and performance of machinery. Forestry and agricultural robots share some challenges and requirements [17], such as the locomotion in rough terrains, localization and mapping in unstructured environments, and planning under uncertainty [18].

A comprehensive fire prevention solution is being developed in the SEMFIRE Project [19], which proposes a multi-robot system to reduce the fuel accumulation in forests and assist in landscaping maintenance. This system consists of small flying robots for vegetation mapping and large-sized tracked mobile robots for forestry mulching.

#### *3.2. Surveillance*

Fire surveillance is the most covered activity in the literature about robotics for firefighting. Most of the proposals involve the use of different kinds of aerial robots (fixed-wing and multi-rotor drones) equipped with various types of cameras (RGB, infrared, multispectral...) to watch over the forests from above. Fire surveillance tasks may have up to four objectives: search of potential fires, detection to alert firefighters, diagnosis to get relevant data about the fire, and prognosis to predict fire propagation [20]. The early detection of fire is as important as the complete analysis of it, given that firefighting teams need information such as the ignition and danger potential to organize their operations [21].

Unmanned aerial vehicles (UAVs) with on-board vision systems have considerable potential in the detection and monitoring of forest fires, since they offer high maneuverability, flexible perspective and resolution, and limited risks to people [22]. For this purpose, surveillance systems should integrate six elements: a fleet of UAVs with payloads, sensor fusion and image processing methods, guidance, navigation and control (GNC) algorithms, coordination and cooperation strategies, path planning algorithms, and ground control stations (GCS) [23]. The selected UAVs shall meet a set of requirements, such as long flight time, accurate localization with the data obtained by the Inertial Measurement Unit (IMU) and Global Navigation Satellite System (GNSS), stable and robust flight, and good image quality [24].

There are multiple approaches to develop vision systems to detect fires. The work in [25] comprehensively analyzes the potential sensors and methods for terrestrial, aerial, and satellite-based fire detection systems. Regarding the hardware, they use visible [26,27], thermal [28,29], multispectral [30,31] and infrared cameras [20,32], as well as environmental sensors (mostly used in indoor scenarios [33], but also proposed for forests [21]). Regarding the software, traditional computer vision algorithms [22,34] compete with recent artificial intelligence solutions [35,36]. The most common features used to recognize fires in aerial images are color, geometry, and movement. Color and geometry allow detecting potential fires in isolated frames, whereas movement is relevant to check these detections with the whole sequence of frames [22]. A challenge for these algorithms is adapting to different types of fires and scenarios: for instance, subterranean fires show up as columns of smoke, in contrast to common surface forest fires [37].

Heterogeneous multi-robot systems are also considered for fire surveillance. The work presented in [38] proposes an air-ground robotic team, where the Unmanned Ground Vehicles (UGVs) compensate for the weaknesses of UAVs, such as their limitations in autonomy (flight time) and payload (weight capacity). This work proposes the use of UGVs to transport UAVs to the fire scenario, where UAVs can take off, perform their tasks, and land again. Additionally, UGVs are used as base stations for UAVs, centralizing the communications between the fleet, processing the data collected by them, and coordinate their tasks in the scenario. Moreover, the work published in [35] proposes the use of two different types of drones: fixed-wing UAVs for medium-altitude flights searching fires and rotary-wing UAVs for low-altitude flights checking detections. The need for checking detection to avoid false alarms is also expressed in [39], which suggests the use of multiple drones to collect simultaneous information of every area, as well as the use of various features to detect fires in the provided images (e.g., color and movement).

#### *3.3. Extinguishing*

In general terms, fire extinguishing is the last task of firefighting after having detected and checked the fire. Currently, this task is mostly performed with ground and aerial means that require human intervention. Sometimes, the presence of humans results in risky situations for their lives due to the virulence of fires. Although this is a good reason to try to use robots in these tasks, these autonomous systems are only used experimentally. The literature considers two main approaches: one for aerial extinguishing and another for supporting ground operations.

The main idea of the firefighting drones is to attack the fire when it is in its first stages, trying to avoid the spread of it. An extinguishing quadcopter equipped with a bucket to capture and release water is presented in [40]. Although this design is similar to those used in current firefighting helicopters, the limitation in the payload capacity of the quadcopter reduces its performance.

An aerial hose-type robot that can fly directly into the fire source by a water jet is presented in [41]. This robot receives a continuous intake of water for fighting the fires and controlling its stability. In this way, it solves the limited payload issues of conventional drones, but it requires a water source close to the fire scenario. Another alternative is the use of gases instead of water. A quadcopter that carries a balloon filled with helium is proposed by [42]. This inert gas is used because it can reduce the amount of oxygen of the flames, as well as it is light enough to be transported by a quadcopter. The scalability of this system to forest fires must be validated, including the mechanism to release the balloons on the exact points.

A common idea for putting out fires is the utilization of extinguishing balls [43]. These elements burst when they come into contact with high temperatures, releasing some chemical components that put out the fire. Ref. [44] proposes a quadcopter that can launch an extinguishing ball to the flames of urban and wildfires. Following the same approach, Refs [45,46] propose some alternatives for the release mechanism to allow throwing multiple balls and keep the stability of the drone. Finally, ref. [47] poses a

swarm of UAVs that can perform monitoring and extinguishing tasks, demonstrating the scalability of fire extinguishing systems based on drones that release balls.

Different types of multi-robot systems are proposed for fire extinguishing missions. There is a trend in the literature to apply multiple light robots instead of developing drones with the capabilities of planes and helicopters. For instance, a drone fleet is proposed in [48] and a drone swarm in [49]. When multiple drones work in the same scenario, the coordination of the fleet becomes relevant. The literature contains various proposals of algorithms to allocate targets among the drones, seeking to minimize traveling distance for every drone. Some examples are [11], which proposes that the team shares all the information of the mission and runs an auction-based mechanism to distribute the tasks, and [50], which describes a deep learning method to allocate tasks, overcoming the sensing, communication, and motion limitations of drones.

In addition to fire extinguishing tasks, robots can be used to monitor fires and provide information to firefighters. Ref. [51] describes a novel algorithm for safe human-robot coordination in wildfires. The drones track the evolution of fires, which can be stationary, moving, and moving/spreading, and a human safety module detects if there are humans close to fire spots. Moreover, ref [52] three types of drones to perform patrolling, confirmation, and monitoring tasks, as well as a fire-spreading model to use the information collected from the fires to predict their behavior.
