A Retrospective Analysis of Indoor CO2 Measurements Obtained with a Mobile Robot during the COVID-19 Pandemic

This work presents a retrospective analysis of indoor CO2 measurements obtained with a mobile robot in an educational building after the COVID-19 lockdown (May 2021), at a time when public activities resumed with mandatory local pandemic restrictions. The robot-based CO2 measurement system was assessed as an alternative to the deployment of a net of sensors in a building in the pandemic period, in which there was a global stock outage of CO2 sensors. The analysis of the obtained measurements confirms that a mobile system can be used to obtain interpretable information on the CO2 levels inside the rooms of a building during a pandemic outbreak.


Introduction
The concept of indoor air quality is generally associated with the health and comfort of a building's occupants [1].Carbon dioxide (or CO 2 ) is a chemical compound, produced as a byproduct of the breathing process, that must be analyzed when assessing air quality [2] due to its negative effects on human health, particularly when present in high concentrations [3][4][5][6][7][8].On average, humans produce around 100 times the original quantity of inhaled CO 2 [3], a high exchange ratio that can contribute to a noticeable rise in CO 2 concentrations, especially in crowded or poorly ventilated indoor spaces.This highlights the importance of keeping an adequate ventilation regime in shared indoor spaces, which is not a trivial problem.In this context, Ding et al. [8] reviewed the classroom ventilation regimes of several educational buildings, concluding that unsupervised ventilation strategies are not sufficient to guarantee a healthy environment for students.
The concentration of CO 2 measured inside a room can be used as a biomarker to estimate if the air inside the room is adequately renewed to prevent the build-up of undesired gases [9,10].Di Gilio et al. [11] proposed the implementation of specific threshold levels to evaluate the risk of inhaling the air exhaled by others in a poorly ventilated room, which should be used as a trigger mechanism to set corrective actions.
In relation to the global pandemic of coronavirus disease 2019 (COVID-19) that was caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the scientific literature [12][13][14] highlighted that the most common infection pathway was the inhalation of contagious airborne particles in the form of small droplets or aerosols [15], which are ejected from the mouth and nose of the infected person via the respiratory system while breathing, talking, or sneezing [16].Zhang et al. [12] analyzed the potential infection risks of people who were exposed to the SARS-CoV-2 virus in public places and non-healthcare environments, concluding that inhalation was the predominant route of exposure, emphasizing the importance of protecting individuals from the airborne transmission of infectious diseases.
In the rise of the COVID-19 pandemic, the World Health Organization (WHO) [17] highlighted that "the risk of getting COVID-19 is higher in crowded and inadequately ventilated spaces where infected people spend long periods of time together in close Sensors 2024, 24, 3102 3 of 16 mainly focused on the use of low-cost [40], battery-powered sensing devices [41].As an example of an IoT application, Marques et al. [39] proposed a custom air quality sensor based on CO 2 analysis that could be deployed in large buildings.The measurements gathered by the device can be accessed through a web or smartphone application.In a similar direction, Vanhaeverbeke et al. [42] proposed a real-time IoT software architecture to calculate and visualize the estimated risk of COVID-19 aerosol transmission in office buildings, using colored lights placed in each room.Their platform was able to process large volumes of data gathered by hundreds of sensors in real-time.
The deployment of a net of stationary CO 2 sensors in an existing building has some drawbacks [43] that should not be underestimated, such as cost, scalability, deployment time, and maintenance.For example, the medium-sized educational building in which the experimental part of this paper was performed would require the deployment of a net composed of 95 stationary CO 2 sensing devices to monitor all its shared spaces.In this context, the use of mobile robots carrying high-quality CO 2 sensors can be an efficient alternative to monitor air quality, especially when taking into consideration the time and cost associated with the deployment of the measurement system.

CO 2 Measurement with Mobile Platforms
The measurement of CO 2 with a mobile platform has been addressed by several authors [44][45][46][47][48]. Jin et al. [44] proposed the development of an automated mobile sensing platform for high-granularity indoor environmental quality monitoring.Quintana et al. [45] proposed using several sensors embedded in a small, low-cost, remotely controlled mobile robot for workplace occupant comfort monitoring.Rahmaniar et al. [46] proposed a hybrid approach, in which a mobile robot was used to carry an air quality sensor connected wirelessly to a base station.Yang et al. [47] proposed the use of a mobile robot to detect and locate a CO 2 contaminant source, obtaining better localization accuracy than when using a net with nine stationary sensors.Curea et al. [48] proposed a semi-autonomous omnidirectional mobile robot to monitor a wide range of ambient parameters, including CO 2 and other harmful industrial gases.
In general, the use of a mobile platform for indoor CO 2 monitoring has the disadvantage that only one location can be analyzed at a time.However, the use of an autonomous mobile robot with an embedded CO 2 sensor also has some advantages: the granularity of the measurements is higher than when using stationary sensors; the calibration and maintenance is simpler; the onboard computational power of the mobile robot can be used to estimate other parameters such as the occupancy of the room and type of activity; the patrolling route can be optimized depending on the planned room occupancy; and CO 2 monitoring can be performed as a secondary task.
The advances of robotic technologies during the COVID-19 pandemic have been analyzed by several authors [34,49].Specifically, Wang et al. [34] reviewed 280 scientific works and discussed 146 of them in detail to identify representative technologies and their readiness level.There were no reports on robot-based CO2 measurement systems used as a preventive tool to fight the transmission of the COVID-19 pandemic.As stated before, this gap inspired the revision of the measurements obtained with a mobile robot embedding a CO 2 sensor during the COVID-19 pandemic.

Materials and Methods
This section describes the materials used in this work: the mobile robot used to carry a CO 2 sensor and perform the measurements, the selected CO 2 sensor, and the methods used for map creation, path-planning, self-localization, path-tracking, and robot-to-building integration.

Mobile Robot
Figure 1 shows the mobile robot used in this work, which was developed by the Automation and Robotics Laboratory of the University of Lleida, Spain [50].This omnidi-rectional humanoid-like robot (1.76 m, 30 kg) was initially designed to provide assistive services [51] and interact with people [52].The robot embeds a Hokuyo UTM-30LX highperformance 2D LIDAR (Hokuyo, Osaka, Japan), which is based on an internal rotating laser-ranging sensor, to conduct a planar 2D exploration (or scan) of the area surrounding the sensor.This LIDAR has a radial distance range of up to 40 m, with a precision between 30 and 50 mm, covers a plane of 270 • around the sensor, and provides 1081 points per scan, at a rate of 40 scans/s.The mobile robot processes the information gathered by this 2D LIDAR for simultaneous localization and mapping (SLAM) [53], obstacle avoidance [54,55], and autonomous path-planning and path-tracking.
tomation and Robotics Laboratory of the University of Lleida, Spain [50].This omnidirectional humanoid-like robot (1.76 m, 30 kg) was initially designed to provide assistive services [51] and interact with people [52].The robot embeds a Hokuyo UTM-30LX highperformance 2D LIDAR (Hokuyo, Osaka, Japan), which is based on an internal rotating laser-ranging sensor, to conduct a planar 2D exploration (or scan) of the area surrounding the sensor.This LIDAR has a radial distance range of up to 40 m, with a precision between 30 and 50 mm, covers a plane of 270° around the sensor, and provides 1081 points per scan, at a rate of 40 scans/s.The mobile robot processes the information gathered by this 2D LIDAR for simultaneous localization and mapping (SLAM) [53], obstacle avoidance [54,55], and autonomous path-planning and path-tracking.
The robot (Figure 1) has a tactile screen for human interaction [52], several onboard cameras, redundant distance sensors, passive infrared detectors (PIRs) for detecting occupancy, and two articulated arms with figurative fixed hands.In this work, the mobile robot embeds a CO2 sensor at a height of 1.3 m above floor level.Preliminary tests showed that placing the sensor at different heights in the structure of the robot did not influence the registered CO2 levels. .

CO2 Sensor Embedded in the Mobile Robot
There are different sensing technologies that can be used to measure CO2 concentrations under indoor conditions [35].The sensor embedded in the mobile robot is based on non-dispersive infrared (NDIR) technology.The non-dispersive infrared (NDIR) CO2 sensor is the S8-LP (Senseair, Stationsgatan 12, 824 71 Delsbo, Sweden), which has an accuracy of ±40 ppm in the measurement ranges from 400 to 2000 ppm and ±3% of reading in ranges from 0 to 10.000 ppm.This sensor requires a warm-up time of at least 30 s after a cold start and measures CO2 concentration, temperature, and humidity once every second.This sampling time was considered adequate for mobile CO2 monitoring.

Method for 2D Map Creation
The deployment of a mobile robot in a building requires a map for planning the exploratory missions, calculating the optimal path to fulfill a mission, self-estimating its relative position and orientation, and tracking the planned path.
The map used by the robot is a 2D point cloud map of the building.In this case, the 2D point cloud map was obtained in a supervised exploration of the building performed before starting the measurements, during a public holiday with no people in the building.The creation of a map requires the following steps: a complete exploration of the building while recording the scans gathered by the 2D LIDAR and the odometry data, automatic registration and merging of the 2D scans in a single 2D map, manual refining of the map, The robot (Figure 1) has a tactile screen for human interaction [52], several onboard cameras, redundant distance sensors, passive infrared detectors (PIRs) for detecting occupancy, and two articulated arms with figurative fixed hands.In this work, the mobile robot embeds a CO 2 sensor at a height of 1.3 m above floor level.Preliminary tests showed that placing the sensor at different heights in the structure of the robot did not influence the registered CO 2 levels.

CO 2 Sensor Embedded in the Mobile Robot
There are different sensing technologies that can be used to measure CO 2 concentrations under indoor conditions [35].The sensor embedded in the mobile robot is based on non-dispersive infrared (NDIR) technology.The non-dispersive infrared (NDIR) CO 2 sensor is the S8-LP (Senseair, Stationsgatan 12, 824 71 Delsbo, Sweden), which has an accuracy of ±40 ppm in the measurement ranges from 400 to 2000 ppm and ±3% of reading in ranges from 0 to 10,000 ppm.This sensor requires a warm-up time of at least 30 s after a cold start and measures CO 2 concentration, temperature, and humidity once every second.This sampling time was considered adequate for mobile CO 2 monitoring.

Method for 2D Map Creation
The deployment of a mobile robot in a building requires a map for planning the exploratory missions, calculating the optimal path to fulfill a mission, self-estimating its relative position and orientation, and tracking the planned path.
The map used by the robot is a 2D point cloud map of the building.In this case, the 2D point cloud map was obtained in a supervised exploration of the building performed before starting the measurements, during a public holiday with no people in the building.The creation of a map requires the following steps: a complete exploration of the building while recording the scans gathered by the 2D LIDAR and the odometry data, automatic registration and merging of the 2D scans in a single 2D map, manual refining of the map, automatic creation of the navigation tree of the map, and manual definition of the reference waypoints used to define the exploratory missions.During this supervised exploration of the building, the robot was manually guided (see Figure 2) [51] to ensure that all educational classrooms were explored.This map only needs to be created once and can be used by all mobile robots operating in the building [56].
Sensors 2024, 24, x FOR PEER REVIEW 5 of 16 automatic creation of the navigation tree of the map, and manual definition of the reference waypoints used to define the exploratory missions.During this supervised exploration of the building, the robot was manually guided (see Figure 2) [51] to ensure that all educational classrooms were explored.This map only needs to be created once and can be used by all mobile robots operating in the building [56].The final 2D point cloud map is created by joining the 2D scans obtained during the exploration by using a variant of the Iterative Closest Point (ICP) algorithm [54,57].The result of this matching process is the determination of a common reference coordinate system for all the scans, allowing them to be unified in a single 2D map. Figure 3 shows the map and the exterior of the educational building explored in this work, which is the Polytechnic School of the Universitat de Lleida, Spain.At this point, the consistency of the map created from a 2D LIDAR depends on the physical characteristics of the walls and furniture elements being scanned [58].In general, the raw map of a facility requires some manual refining to create enclosed areas with the use of virtual walls.This feature is essential, since some obstacles, such as glass panels, hollow handrails, or stairwells, may not be detectable with a 2D LIDAR.Virtual walls created during the manual refinement of the map can also serve to restrict the navigable areas that are accessed by the robot.In this work, the navigable areas are defined by a grid with nodes measuring 0.120 m × 0.120 m each, which is created using the Breadth-First Search (BFS) algorithm [59].The final 2D point cloud map is created by joining the 2D scans obtained during the exploration by using a variant of the Iterative Closest Point (ICP) algorithm [54,57].The result of this matching process is the determination of a common reference coordinate system for all the scans, allowing them to be unified in a single 2D map. Figure 3 shows the map and the exterior of the educational building explored in this work, which is the Polytechnic School of the Universitat de Lleida, Spain.At this point, the consistency of the map created from a 2D LIDAR depends on the physical characteristics of the walls and furniture elements being scanned [58].In general, the raw map of a facility requires some manual refining to create enclosed areas with the use of virtual walls.This feature is essential, since some obstacles, such as glass panels, hollow handrails, or stairwells, may not be detectable with a 2D LIDAR.Virtual walls created during the manual refinement of the map can also serve to restrict the navigable areas that are accessed by the robot.
Sensors 2024, 24, x FOR PEER REVIEW 5 of 16 automatic creation of the navigation tree of the map, and manual definition of the reference waypoints used to define the exploratory missions.During this supervised exploration of the building, the robot was manually guided (see Figure 2) [51] to ensure that all educational classrooms were explored.This map only needs to be created once and can be used by all mobile robots operating in the building [56].The final 2D point cloud map is created by joining the 2D scans obtained during the exploration by using a variant of the Iterative Closest Point (ICP) algorithm [54,57].The result of this matching process is the determination of a common reference coordinate system for all the scans, allowing them to be unified in a single 2D map. Figure 3 shows the map and the exterior of the educational building explored in this work, which is the Polytechnic School of the Universitat de Lleida, Spain.At this point, the consistency of the map created from a 2D LIDAR depends on the physical characteristics of the walls and furniture elements being scanned [58].In general, the raw map of a facility requires some manual refining to create enclosed areas with the use of virtual walls.This feature is essential, since some obstacles, such as glass panels, hollow handrails, or stairwells, may not be detectable with a 2D LIDAR.Virtual walls created during the manual refinement of the map can also serve to restrict the navigable areas that are accessed by the robot.In this work, the navigable areas are defined by a grid with nodes measuring 0.120 m × 0.120 m each, which is created using the Breadth-First Search (BFS) algorithm [59].In this work, the navigable areas are defined by a grid with nodes measuring 0.120 m × 0.120 m each, which is created using the Breadth-First Search (BFS) algorithm [59].This algorithm initiates at the robot's initial coordinates at the start of the exploration and iteratively expands to neighboring nodes until all accessible areas are mapped, avoiding contact with the mapped wall points.Then, each node of the grid is weighted according to its distance to the nearest obstacle or wall.The weights of the nodes are represented in Figure 3 using a color scale: the blue color represents the safest navigable area, while the yellow color represents areas that, although navigable, are much closer to obstacles.
The navigable areas are used to create the navigation tree of the map, which represents the list of interconnections between all its nodes.The granularity of the grid and the connectivity of the nodes define the computational challenge that represents the use of deterministic search algorithms [60] to plan the trajectory of the robot.However, recent works have demonstrated that similar navigation results can be obtained when using a navigation tree that is not based on a discrete grid definition [61].
Finally, the map must be edited to label the nodes depicting relevant destinations (points of interest) such as charging stations, classrooms, elevators, etc.The use of labeled nodes simplifies the process of defining routes and managing the connections between important locations, especially in the case of elevators, which are used as transition points between different floors.

Method for Path-Planning
The path-planning algorithm is responsible for computing a sorted list of nodes to be visited in order to reach the destination defined in a mission [62].The path-planning method used in this work to explore the navigation tree of the building is the A* (A-star) algorithm [60,62,63].Starting from a known waypoint (or node position) on the map, the A* algorithm determines which sequence of nodes in the navigation tree are to be visited to complete an exploratory mission.This sequence includes the possibility of accessing elevators to navigate between the floors of the building during a mission [64].Figure 4 depicts an example of the planned path from the current robot position (violet ring) to a destination defined by a key waypoint (black ring).Figure 4 shows small colored circles detailing the localization of the nodes of the navigation tree in the area.The direct result of the A* algorithm (Figure 4a) initially forces the robot to move in node-sized steps to visit one of its eight neighbors.This result is then smoothed to reduce the total number of nodes to be visited (Figure 4b) [65] while ensuring the safety of the robot [63].
Sensors 2024, 24, x FOR PEER REVIEW 6 of 16 This algorithm initiates at the robot's initial coordinates at the start of the exploration and iteratively expands to neighboring nodes until all accessible areas are mapped, avoiding contact with the mapped wall points.Then, each node of the grid is weighted according to its distance to the nearest obstacle or wall.The weights of the nodes are represented in Figure 3 using a color scale: the blue color represents the safest navigable area, while the yellow color represents areas that, although navigable, are much closer to obstacles.
The navigable areas are used to create the navigation tree of the map, which represents the list of interconnections between all its nodes.The granularity of the grid and the connectivity of the nodes define the computational challenge that represents the use of deterministic search algorithms [60] to plan the trajectory of the robot.However, recent works have demonstrated that similar navigation results can be obtained when using a navigation tree that is not based on a discrete grid definition [61].
Finally, the map must be edited to label the nodes depicting relevant destinations (points of interest) such as charging stations, classrooms, elevators, etc.The use of labeled nodes simplifies the process of defining routes and managing the connections between important locations, especially in the case of elevators, which are used as transition points between different floors.

Method for Path-Planning
The path-planning algorithm is responsible for computing a sorted list of nodes to be visited in order to reach the destination defined in a mission [62].The path-planning method used in this work to explore the navigation tree of the building is the A* (A-star) algorithm [60,62,63].Starting from a known waypoint (or node position) on the map, the A* algorithm determines which sequence of nodes in the navigation tree are to be visited to complete an exploratory mission.This sequence includes the possibility of accessing elevators to navigate between the floors of the building during a mission [64].Figure 4 depicts an example of the planned path from the current robot position (violet ring) to a destination defined by a key waypoint (black ring).Figure 4 shows small colored circles detailing the localization of the nodes of the navigation tree in the area.The direct result of the A* algorithm (Figure 4a) initially forces the robot to move in node-sized steps to visit one of its eight neighbors.This result is then smoothed to reduce the total number of nodes to be visited (Figure 4b) [65] while ensuring the safety of the robot [63].The robot has been designed to operate in dynamic environments with ongoing human activity, so it is expected that new static and dynamic obstacles will be encountered during exploratory missions.As an application example, Figure 4c represents the detection of an obstacle (black dots) in the planned path, which triggers the computation of a The robot has been designed to operate in dynamic environments with ongoing human activity, so it is expected that new static and dynamic obstacles will be encountered during exploratory missions.As an application example, Figure 4c represents the detection of an obstacle (black dots) in the planned path, which triggers the computation of a new trajectory to avoid the obstacle.In this case, the onboard computational power of the robot enables real-time re-evaluation of the path by executing the A* algorithm on the spot to automatically update the planned path upon detecting an obstacle.

Methods for Self-Localization and Path-Tracking
The method employed for robot self-localization relies on matching the current 2D scan provided by its 2D LIDAR sensor with the 2D map of the building floor by using the ICP algorithm [54,57].This matching provides an estimate of the relative location and orientation of the robot on the 2D map.The obstacle detection procedure implemented in the robot is performed after the ICP matching [54,66].The points of the scan retrieved by the LIDAR sensor which have not been matched with other points of the map are automatically identified as obstacles to be avoided.
The path-tracking procedure implemented in the robot is described in [67].This procedure is based on sampling the planned path to define a sequence of intermediate trajectory waypoints to be visited by the omnidirectional robot.

Method for Robot-to-Building Integration
At the time when this work was performed, it was not possible to interact or send information to the mechanical ventilation system of the building.As an alternative, the robot was configured to send warning messages to a remote operator when one of the monitored parameters exceeded a predefined threshold.Therefore, if necessary, the alerted remote operator would re-adjust the heating, ventilation, and air conditioning (HVAC) system of the building to implement corrective measures.The technical challenges and benefits of an interactive relationship between a mobile sensing platform and the existing demand-driven HVAC [68] of a building will be explored in future works.

Exploratory Missions
Table 1 summarizes the exploratory missions scheduled to register the CO 2 levels in five rooms of an educational building.Classrooms C0.01, C0.02, C0.03, C0.04, and C0.05 (defined in Figure 3) have a similar size, layout, furniture distribution, entrance door, and windows.The classrooms are designed to carry out quiet lectures, an educational activity involving no physical activity, and in which only the lecturer is usually speaking or moving in front of a blackboard.For the purpose of mobile robot exploration, each classroom is identified by a node-label or waypoint, which is physically defined not too far into the classroom to minimize possible lecture disruptions while performing CO 2 measurements.1 defines two different missions.The first mission M0, scheduled at 8:00 h, was proposed to validate the correct operation of the robot before starting the educational activities in the building and to record reference or background CO 2 levels.The following missions, from M1 to M4, were scheduled throughout the morning to register the CO 2 level inside the classrooms in accordance with the planned occupancy.In all these missions, the robot started from a charging station, tracked the planned trajectory, stopped after reaching the coordinates of the designated point of interest inside the different classrooms, waited for 5 min in order to obtain a stable CO 2 measurement, and returned to the charging station.During the exploratory missions, the mobile robot was instructed to move at a medium-low translational velocity (0.2 m/s) and configured in silent mode with no sound generation.

CO 2 Measurement Results
This section presents the CO 2 levels obtained by the mobile platform while performing the exploratory missions M0 to M4.The measurements were carried out in May of 2021, after a local COVID-19 lockdown, when activities were gradually resumed with the compulsory pandemic restrictions applied [22,69].Specifically, these were as follows: windows and doors open, 10 min break after 50 min of activity, students seated at least one meter away from each other, and compulsory use of face masks.In such pandemic conditions, all the doors of the classrooms that the robot encountered on its exploration path were open, and using a mobile robot to enter classrooms to monitor the CO 2 level was perceived as safer than if this operation were performed by a human operator.
Figure 5 shows the georeferenced evolution of the CO 2 levels registered during the M0 mission.The CO 2 levels registered ranged from 404 ppm in the hallway to 468 ppm inside C0.05 classroom.Similarly, Figure 6 shows the georeferenced evolution of the CO 2 levels registered during the M1, M2, M3, and M4 missions.Figure 6 illustrates the presence of inertia in the measurements obtained from the CO 2 sensor, but the 5 min waiting period inside the classrooms was deemed sufficient for obtaining valid measurements.
after reaching the coordinates of the designated point of interest inside the different classrooms, waited for 5 min in order to obtain a stable CO2 measurement, and returned to the charging station.During the exploratory missions, the mobile robot was instructed to move at a medium-low translational velocity (0.2 m/s) and configured in silent mode with no sound generation.

CO2 Measurement Results
This section presents the CO2 levels obtained by the mobile platform while performing the exploratory missions M0 to M4.The measurements were carried out in May of 2021, after a local COVID-19 lockdown, when activities were gradually resumed with the compulsory pandemic restrictions applied [22,69].Specifically, these were as follows: windows and doors open, 10 min break after 50 min of activity, students seated at least one meter away from each other, and compulsory use of face masks.In such pandemic conditions, all the doors of the classrooms that the robot encountered on its exploration path were open, and using a mobile robot to enter classrooms to monitor the CO2 level was perceived as safer than if this operation were performed by a human operator.
Figure 5 shows the georeferenced evolution of the CO2 levels registered during the M0 mission.The CO2 levels registered ranged from 404 ppm in the hallway to 468 ppm inside C0.05 classroom.Similarly, Figure 6 shows the georeferenced evolution of the CO2 levels registered during the M1, M2, M3, and M4 missions.Figure 6 illustrates the presence of inertia in the measurements obtained from the CO2 sensor, but the 5 min waiting period inside the classrooms was deemed sufficient for obtaining valid measurements.The outcomes of M1, M2, M3, and M4 missions are also presented in Figure 7, employing a time scale evolution to evince that the mobile robot provides sampled information on the CO2 levels inside the explored classrooms.Figure 7a shows the CO2 levels obtained inside C0.01 classroom during two exploratory missions (M1 and M2) and the number of people attending the lecture.During the 5 min wait, the concentration increases after reaching the coordinates of the designated point of interest inside the different classrooms, waited for 5 min in order to obtain a stable CO2 measurement, and returned to the charging station.During the exploratory missions, the mobile robot was instructed to move at a medium-low translational velocity (0.2 m/s) and configured in silent mode with no sound generation.

CO2 Measurement Results
This section presents the CO2 levels obtained by the mobile platform while performing the exploratory missions M0 to M4.The measurements were carried out in May of 2021, after a local COVID-19 lockdown, when activities were gradually resumed with the compulsory pandemic restrictions applied [22,69].Specifically, these were as follows: windows and doors open, 10 min break after 50 min of activity, students seated at least one meter away from each other, and compulsory use of face masks.In such pandemic conditions, all the doors of the classrooms that the robot encountered on its exploration path were open, and using a mobile robot to enter classrooms to monitor the CO2 level was perceived as safer than if this operation were performed by a human operator.
Figure 5 shows the georeferenced evolution of the CO2 levels registered during the M0 mission.The CO2 levels registered ranged from 404 ppm in the hallway to 468 ppm inside C0.05 classroom.Similarly, Figure 6 shows the georeferenced evolution of the CO2 levels registered during the M1, M2, M3, and M4 missions.Figure 6 illustrates the presence of inertia in the measurements obtained from the CO2 sensor, but the 5 min waiting period inside the classrooms was deemed sufficient for obtaining valid measurements.The outcomes of M1, M2, M3, and M4 missions are also presented in Figure 7, employing a time scale evolution to evince that the mobile robot provides sampled information on the CO2 levels inside the explored classrooms.Figure 7a shows the CO2 levels obtained inside C0.01 classroom during two exploratory missions (M1 and M2) and the number of people attending the lecture.During the 5 min wait, the concentration increases The outcomes of M1, M2, M3, and M4 missions are also presented in Figure 7, employing a time scale evolution to evince that the mobile robot provides sampled information on the CO 2 levels inside the explored classrooms.Figure 7a shows the CO 2 levels obtained inside C0.01 classroom during two exploratory missions (M1 and M2) and the number of people attending the lecture.During the 5 min wait, the concentration increases and stabilizes, although the time scale of the figure makes it difficult to observe this stabilization.Figure 7a shows that M2 mission registers a higher CO 2 level inside the room, revealing the cumulated effect of the CO 2 exhaled during the lecture (despite having the windows and door open).Similarly, Figure 7b shows the CO 2 levels obtained inside C0.02 classroom during two exploratory missions (M3 and M4) and the number of people attending the lecture.In these two measurements, the CO 2 concentration stabilized at nearly the same Sensors 2024, 24, 3102 9 of 16 level, probably due to an equilibrium between the exhaled CO 2 during the lecture and the inflow of fresh air into the room.Finally, Figure 7c shows the CO 2 levels obtained inside C0.03 classroom during the four M1 to M4 exploratory missions, which was empty throughout the day.The sensor needs some time to show the change in CO 2 concentration inside the classroom.However, after the 5 min wait, the registered CO 2 level is practically the same in all the missions.and stabilizes, although the time scale of the figure makes it difficult to observe this stabilization.Figure 7a shows that M2 mission registers a higher CO2 level inside the room, revealing the cumulated effect of the CO2 exhaled during the lecture (despite having the windows and door open).Similarly, Figure 7b shows the CO2 levels obtained inside C0.02 classroom during two exploratory missions (M3 and M4) and the number of people attending the lecture.In these two measurements, the CO2 concentration stabilized at nearly the same level, probably due to an equilibrium between the exhaled CO2 during the lecture and the inflow of fresh air into the room.Finally, Figure 7c shows the CO2 levels obtained inside C0.03 classroom during the four M1 to M4 exploratory missions, which was empty throughout the day.The sensor needs some time to show the change in CO2 concentration inside the classroom.However, after the 5 min wait, the registered CO2 level is practically the same in all the missions.

Evaluation of the Risk of COVID-19 Infection
At the time when the exploratory missions were conducted (May 2021), the evaluation of the risk of COVID-19 infection was based on the safe threshold CO2 levels proposed by Di Gilio et al. [11] to reduce the risk of SARS-CoV-2 airborne transmission (shown in

Evaluation of the Risk of COVID-19 Infection
At the time when the exploratory missions were conducted (May 2021), the evaluation of the risk of COVID-19 infection was based on the safe threshold CO 2 levels proposed by Di Gilio et al. [11] to reduce the risk of SARS-CoV-2 airborne transmission (shown in Table 2).These threshold levels can be adapted depending on the pathogen transmitted [11].Table 3 shows the classroom monitored, the mission time, the expected classroom occupancy, the cumulated occupancy time, the maximum CO 2 levels registered by the robot, and an estimation of the risk of COVID-19 infection according to the safe threshold levels applied.The maximum CO 2 levels registered in all the monitored classrooms were below 700 ppm, so the risk of SARS-CoV-2 airborne transmission in the occupied classrooms was always evaluated as "Low Risk".generated inside the classrooms.Table 4 details the maximum CO 2 excess registered inside the classrooms (relative to the reference background CO 2 level of 468 ppm).Table 4 also details the type of activity performed in the classrooms, along with the use of face masks by classroom users, which was 100% due to local regulations that made wearing them compulsory in that period (May 2021) [22].The expected number of infected people (presumably asymptomatic) in the classrooms was obtained from governmental data sources offering public average regional information [73] and was rounded to the upper integer: one person for the occupancy expected in the classrooms.Similarly, the percentage of immune people in the classrooms was also obtained from governmental data sources [74,75].In this case, the availability of official data sources offering regional information of the evolution of the COVID-19 pandemic [73][74][75] avoided the task of collecting individual information directly from the people present in the classroom activities.A Compulsory use of face masks.B Obtained from local governmental data sources [73,74].
Table 4 details the value of the safe CO 2 excess limit estimated with the model proposed by Bazant et al. [70].In the case of C0.01 classroom, this value is higher than 1600 ppm for an occupancy time of up to 1 h, dropping to 1325 ppm after 2 h.In the case of C0.02 classroom, this value is 891 ppm for an occupancy time of 3 h.Comparatively, these safe values are significantly higher than the absolute threshold levels defined by Di Gilio et al. [11], since the dynamic model [70] also includes the filtering effect provided by the use of face masks, which significantly increases the safe CO 2 threshold level during an indoor activity [76,77].
Finally, Table 4 provides an estimation of the risk of SARS-CoV-2 aerosol transmission based on the proposal of Vanhaeverbeke et al. [42]: dividing the measured CO 2 excess by the value of the safe CO 2 excess provided by the model [70] and clipping the results in a range from 0 (0% risk) to 1 (100% risk).Therefore, a risk of 100% means that the safe CO 2 excess of the indoor activity has been surpassed, so that a healthy, non-immune occupant of the room is at risk of being infected by airborne transmission of SARS-CoV-2.The risks that were finally assessed were 0% for an occupancy time up to 1 h, rising to 15% after 2 h and to 25% after 3 h.These risk values confirm the adequacy of the measures deployed to reduce the risk of airborne transmission of the SARS-CoV-2 virus [22].

Discussion and Conclusions
This paper presents the work carried out with an autonomous robot-based measurement system embedding a CO 2 sensor during the COVID-19 pandemic.This work was developed at a time when there was a global stock outage of CO 2 sensors [32].The hypothesis was that a mobile measurement system can be used as an alternative to the deployment of a net of sensors to measure the CO 2 levels of different rooms in a building.
Table 5 compares the mobile measurement system deployed in this work with the referenced robots embedding a CO 2 sensor [44][45][46][47][48].The main differential elements of the proposed robot are its size (1.76 m, 30 kg), its human interaction capability [52], and its A series of experiments were conducted in an educational building to measure the CO 2 levels in different classrooms with the robot.These experiments were performed in May of 2021 after a local COVID-19 lockdown, when activities resumed following the mandatory pandemic restrictions established by the government [22,69].During the experiments, all the doors of the classrooms that the robot encountered on its exploration path were open, and the use of a mobile robot to monitor the CO 2 levels was perceived as safer compared to human operators performing the same monitoring task.
The CO 2 levels obtained during the explorations were evaluated in real-time according to the risk classification scheme proposed by Di Gilio [11] to reduce the risk of SARS-CoV-2 airborne transmission.All monitored classrooms received the classification of "Low Risk" of SARS-CoV-2 airborne transmission (see Table 3).
This work also presents a retrospective analysis of the CO 2 levels registered.This new analysis is inspired by the work carried out in 2023 by Vanhaeverbeke et al. [42], which evaluated the risk of COVID-19 airborne transmission from real-time CO 2 measurements based on the mathematical model proposed by Bazant et al. [70,71].This model provides an upper bound of the CO 2 excess, tailored to prevent aerosol transmission of SARS-CoV-2 in indoor environments, depending on the occupancy and time cumulated in a room.The application of this model showed a maximum risk of SARS-CoV-2 airborne transmission in the classrooms of 25% (see Table 4), a value that confirms the adequacy of the pandemic restrictions applied in the building.
The conclusion of this work is that a mobile measurement system can be used to obtain interpretable information on the CO 2 levels inside the rooms of a building.The development of mobile CO 2 monitoring capabilities is aligned with the advice of the World Health Organization (WHO) [78], highlighting that "future pandemics are inevitable".The use of the robot to estimate the size and occupancy of a room, as well as the use of face masks, will be studied in future works.The measurement system will be adapted to improve the control and sustainability of the management of the HVAC [68,79].Finally, another promising improvement is the integration of a CO 2 sensor in a compact eNose, trained to operate fully as an early non-invasive detector of airborne transmitted pandemic outbreaks [80,81].

Figure 1 .
Figure 1.Mobile robot performing a routine inspection.

Figure 1 .
Figure 1.Mobile robot performing a routine inspection.

Figure 2 .
Figure 2. Images of one of the authors of this work manually guiding the robot while performing the supervised exploration of a building during the COVID-19 pandemic.

Figure 3 .
Figure 3. Two views of the experimentation building: (left) point cloud map of its first floor including classroom numbers and a color scale identifying the navigable areas (colored nodes), and (right) 3D photometric reconstruction of the exterior of the building.

Figure 2 .
Figure 2. Images of one of the authors of this work manually guiding the robot while performing the supervised exploration of a building during the COVID-19 pandemic.

Figure 2 .
Figure 2. Images of one of the authors of this work manually guiding the robot while performing the supervised exploration of a building during the COVID-19 pandemic.

Figure 3 .
Figure 3. Two views of the experimentation building: (left) point cloud map of its first floor including classroom numbers and a color scale identifying the navigable areas (colored nodes), and (right) 3D photometric reconstruction of the exterior of the building.

Figure 3 .
Figure 3. Two views of the experimentation building: (left) point cloud map of its first floor including classroom numbers and a color scale identifying the navigable areas (colored nodes), and (right) 3D photometric reconstruction of the exterior of the building.

Figure 4 .
Figure 4. Path-planning example.The blue dots represent the point cloud of the map, while the green dots represent the current 2D scan points that match with the map: (a) the discrete trajectory found with the A* algorithm to navigate to a key waypoint using the navigation tree with a grid of 0.12 m × 0.12 m; (b) the optimized path avoiding unnecessary intermediate maneuvers; and (c) the recalculated path to avoid an obstacle (black points) detected in the planned path.

Figure 4 .
Figure 4. Path-planning example.The blue dots represent the point cloud of the map, while the green dots represent the current 2D scan points that match with the map: (a) the discrete trajectory found with the A* algorithm to navigate to a key waypoint using the navigation tree with a grid of 0.12 m × 0.12 m; (b) the optimized path avoiding unnecessary intermediate maneuvers; and (c) the recalculated path to avoid an obstacle (black points) detected in the planned path.

Figure 5 .Figure 6 .
Figure 5. Georeferenced representation of the CO2 concentration evolution recorded by the mobile robot during the M0 exploratory mission.

Figure 5 .
Figure 5. Georeferenced representation of the CO 2 concentration evolution recorded by the mobile robot during the M0 exploratory mission.

Figure 5 .Figure 6 .
Figure 5. Georeferenced representation of the CO2 concentration evolution recorded by the mobile robot during the M0 exploratory mission.

Table 2 .
[11] classification scheme proposed by Di Gilio[11]to reduce the risk of SARS-CoV-2 airborne transmission based on CO 2 concentration monitoring.

Table 4 .
[70]mation of the risk of SARS-CoV-2 aerosol transmission based on the CO 2 excess registered and the model of the safe CO 2 excess proposed by Bazant et al.[70].