Mobile robots exploration strategies and requirements: A systematic mapping study

4.1.1. Environment

Robotic exploration tasks are strongly affected by the characteristics of the environment that the robot has to explore. In the proposed taxonomy, we characterize the environment with three classification facets: Structure, Density, and Barriers.

Structure identifies the type of scene in which the robot is performing the exploration task. This can be indoor areas like Office-like and Home environments, which are characterized by several rooms and corridors that are similar to each other. Home environments are generally less structured than office-like areas, due to the higher variety in furniture and layouts. Collapsed buildings are characterized by a cluttered composition due to the presence of elements such as pipes, wires, debris, or even destroyed surfaces. Subterranean environments, both natural like caverns and artificial like underground areas, are very large corridor-like scenarios with repetitive tunnels (Ohradzansky et al., 2020). In the primary studies that we have analyzed, Industrial plants are large outdoor areas characterized by industrial buildings. In City-like scenarios the main elements are residential buildings, streets, cars, and eventually people. Other scenarios include Outdoor fields like in forests or agricultural scenarios, Underwater environments with diverse ecosystems, and Planetary surfaces characterized by dust and rocks.

The Density classification facet characterizes the volume of the scene and the spatial distribution of obstacles. In particular, we describe the environment as Small in the case of a single room and as Large for indoor scenarios that are made of multiple rooms as in Grontas et al. (2020) and for outdoor environments. In addition, an environment is considered Open if it is characterized by large passages compared to the robot size (e.g., an empty room, a hospital ward, a lawn). Otherwise, it is considered Cluttered (e.g., a forest or a crowded room). In the taxonomy, we considered a combination of these two characteristics. Examples of Small Open and Small Cluttered are presented by Wang et al. (2019a) and Naazare et al. (2022), respectively. While Batinovic et al. (2022) and Zhong et al. (2022) proposed exploration strategies in Large Open and Large Cluttered environments.

The environment can be further characterized by the presence of Barriers that can limit communication between robots (Best et al., 2023), affect the accuracy of mapping and localization due to poor illumination, the absence of natural landmarks in the environment (Papachristos et al., 2019b), cause severe odometry drifts (Schmid et al., 2021; Zhang et al., 2022c), or make the navigation more risky due to the presence of dynamic obstacles (Zheng et al., 2022a), hard rough terrain (Lehner et al., 2017), or even stairs (Bouman et al., 2020).

Other barriers are implicitly imposed by the type of environment and are not explicitly represented in the taxonomy. For example, indoor, subterranean, underwater, and planetary surfaces are GPS-denied.

4.1.2. Task

As suggested by Garaffa et al. (2023), the proposed taxonomy distinguishes between goal-guided and non-guided exploration scenarios.

In non-guided scenarios, the task focuses on the exploration of unknown areas of the environment for Area Mapping, that is, building a map that can be later used for navigation, or for more specific purposes. In particular, Terrain Learning consists in using specific sensors (e.g., force sensors or RGB cameras) to acquire soil compact information for traversability analysis (Fentanes et al., 2018). Similarly, Area Monitoring consists of inspecting an entire area and collecting information, such as the salinity level of the water, as in Suresh et al. (2020), and differs from target-searching applications, where the task ends once a target is located.

In goal-guided scenarios, the process of exploring an unknown environment still produces a map, but it is guided by information provided by additional sensors, such as visual information in Search & Rescue (Niroui et al., 2019) and Search for Objects (Goel et al., 2021) operations, or a target position with regards to a reference frame in Target Pursuing (Arvanitakis et al., 2018).

4.1.3. Embodiment

The robot’s kinematics, its mobility mechanisms, and the onboard sensors characterize the robot Embodiment, which is typically designed to take into account the characteristics of the operational environment and exploration task.

In our taxonomy, we consider three categories of mobile robots, that is, Unmanned Ground Vehicle (UGV), Unmanned Aerial Vehicle (UAV), and Unmanned Underwater Vehicle (UUV). According to the analysis of the primary studies, UGVs are the most used platform to perform exploration and include both wheeled robots (Zurbrugg et al., 2022) and legged robots (Bouman et al., 2020). UAVs differ for their dimension and include micro UAVs (Respall et al., 2021). Typically, UUVs have a torpedo shape as in (Palomeras et al., 2019a).

The taxonomy also considers multi-robot systems composed of robots of the same type, that is, UGV-UGV (Smith and Hollinger, 2018), UAV-UAV (Renzaglia et al., 2020), and UUV-UUV (Griffiths Sánchez et al., 2018). In the selected primary studies we have found only examples of hybrid systems composed of one UGV and one UAV like in Kulkarni et al. (2022) and represented in the taxonomy with the classification facet UGV-UAV.

Some approaches documented in the selected primary studies (e.g., Kan et al. (2020)) are based on the use of different sensors for navigation and mapping and for collecting observations from the environment as in terrain learning applications (e.g., Papachristos et al. (2019b)).

In the proposed taxonomy, the sensors for navigation and mapping are Sonar, 2D Laser and 3D Laser scanners, Stereo Camera, Structured Light Camera, and Time-of-Flight ToF Camera. Some authors generically indicate that their approach uses depth images without specifying the type of sensor that generates them. In this case, we use the category Depth Camera. Additional sensors for acquiring semantic information are Thermal Camera, RGB Camera, Infrared Sensor, Force Sensor, and RFID Reader.

Exploration strategies strongly depend on the set of onboard sensors, which are chosen taking into account the environmental conditions and other application requirements. For example, in case of poor illumination, a stereo camera does not provide accurate and reliable information, while a laser device can retrieve the 3D shape of the environment even in such conditions. Another example can be the use of thermal cameras that is generally linked to applications like search and rescue operations, to be able to identify more easily the person to be rescued (Doroodgar et al., 2014).

4.2.1. Map type

The main outcome of an exploration task is the map of the environment, which represents the robot’s knowledge about its workspace and surrounding environment as well as objects contained therein (Kucner et al., 2023).

A review of the various types of maps is beyond the scope of this SMS and highly cited secondary studies can be found in the literature, such as Thrun (2003). In Section 5, we will discuss the recurrent associations found in selected primary studies between each type of map, the exploration strategies, and the application requirements.

Here, for the sake of clarity, we briefly recall the common definitions of the Map Types that have been used for the exploration approaches documented in the selected primary studies.

2D/3D Occupancy maps employ a tessellation of the environment into cells, that is, a discrete regular decomposition, where each cell stores the probability that a corresponding area or volume is occupied. The state of a cell is defined as a discrete random variable with two values, occupied and empty (Elfes, 1989).

More general models of occupancy maps have been proposed to overcome the limitations of the original discrete representation, such as the inability to capture structural correlations present in the environment.

As illustrated by O’Callaghan and Ramos (2012) and Kim and Kim (2015), the Gaussian Process occupancy map is a continuous occupancy representation of the environment, which overcomes the assumption of independence between grid cells. This type of map has been used by Ghaffari Jadidi et al. (2019) for computing frontier cells efficiently using the gradient of the Gaussian Process occupancy map.

Hilbert maps (Ramos and Ott, 2016) are a continuous version of occupancy maps that capture statistical relationships between sensory measurements, thus being more robust to outliers. One significant advantage of Hilbert maps for robot exploration is the compression capabilities of the model, which can reduce the cost of communicating maps between the members of a multi-robot team. Francis et al. (2019) present an example of an exploration strategy based on Hilbert maps.

Truncated Signed Distance Field (TSDF) is a volumetric representation of a scene that was originally proposed as a method for integrating range images by Curless and Levoy (1996). It is similar to 3D occupancy grids, where each voxel stores the distance to the nearest surface (Werner et al., 2014). It is well suited for incremental updating and for the implementation of data-parallel algorithms on GPUs. We have found a significant number of recent primary studies that use this type of representation, such as Schmid et al. (2021) and Zhang et al. (2022c).

A more recent method for 3D mapping is represented by Neural implicit representations (Lionar et al., 2021), which implicitly represents the scene by means of a feedforward artificial neural network, where shapes and appearance are encoded in the weights of the neural network. One main advantage of Neural implicit representations is the ability to decouple the memory cost of the representation from the actual spatial resolution. This means that a fixed number of parameters can encode a surface with arbitrarily fine resolution.

Elevation map, generally addressed as a 2.5D map, represents the environment as a two-dimensional surface that is discretized in cells, each one encapsulating the elevation of the corresponding area of the terrain (Hebert et al., 1989).

3DMesh, or polygon mesh, is a group of vertices, edges, and faces that represent the 3D shape of a polyhedral object (Lorensen and Cline, 1987).

Topological map is a symbolic representation of the space that comprises a set of vertices describing a certain area and the involved edges, connecting the vertices to indicate connectivity (Thrun, 1998).

Sematic Maps enrich spatial representation of the environment with semantic labels (Maturana, 2021). As an example, the exploration strategy proposed by Asgharivaskasi and Atanasov (2023) builds a 3D occupancy grid, where each cell is labeled with an environmental category, such as building, vegetation, and terrain. In this case, and similar cases, we classify the approach with both categories, that is, Sematic Map and 3D Occupancy. Similarly, the approach proposed by Zurbrugg et al. (2022) extends TSDF maps with semantic information related to the objects typically found in a home environment: floor, bed, table, door, etc.

4.2.2. Planning the next exploration step

As described in Section 1, the central aspect of an exploration strategy is the identification at each iteration of the best accessible pose in the explored area represented in the current map, where the robot is supposed to observe new portions of the environment and gather new information. Each strategy defines what “best” means and exploits specific techniques for planning the next exploration step.

In theory, all the physically reachable destinations in the current map should be evaluated, as they may reveal new information about the unknown part of the environment. However, in real applications, the brute-force approach is not feasible due to the computational complexity that grows exponentially with the search space (Lluvia et al., 2021). For this reason, alternative definitions of the search space have been proposed. Instead of optimizing over the space of reachable locations, some approaches optimize over actuator controls or trajectories. The advantage of these approaches is the possibility to optimize the exploration activity taking into account additional constraints explicitly, such as total exploration time, available energy, and communication distance.

As proposed by Garaffa et al. (2023), exploration strategies can be classified into two categories: two-stages approaches and end-to-end approaches.

Two-stages approaches perform two steps in sequence: Viewpoint Generation and Viewpoint Selection.

Viewpoint Generation (see the list of papers in Table 8) consists of generating an adequate number of candidate poses where the next perception action can be performed (searching the space of locations, controls, or trajectories) with a sort of filter that reduces the size of the search space. In our taxonomy, we consider three families of viewpoint generation approaches: those based on partitioning the free area of the map, those that are based on the random generation of candidates, and those based on the identification of visual features.

The Area partitioning family includes four methods. The Frontier detection method was first proposed by Yamauchi (1997b). It consists in exploiting the information of occupied and unknown areas of the environment to detect the boundary regions among them, called frontiers. The frontiers are most commonly extracted from a geometric map and are grouped in clusters. For each cluster, the algorithm identifies a viewpoint. Some approaches (e.g., Cieslewski et al. (2017)) adopt a more reactive strategy and restrict the considered frontiers to those that are in the current field of view of the onboard sensor. This strategy specifically takes into account the limited power of UAVs avoiding the effort to significantly change the current velocity. Other approaches (e.g., Shen et al. (2012)) consider as frontiers unstructured regions of the map, which are identified by simulating the dispersion of particles through the known and unknown space. This strategy is preferable to the classical frontier-based approach when the environment is very cluttered. Since frontiers identification is a computationally intensive task, some strategies (e.g., Shade and Newman (2011)) adopt optimization methods instead of the classical geometric method.

Voronoi Graphs are used to decompose the space into regions based on the distance between a set of points. This process is also called “tessellation.” These regions are then used to find clear routes in space, extracting collision-free candidate poses (Kim, 2014).

Ergodic trajectories are a recent method applied in robotic exploration that comes from the ergodic theory, a branch of mathematics that focuses on the statistical properties of a deterministic dynamical system. In particular, it refers to the fraction of time spent in the sampling area, which should be equal to some metric (varying with the application) that quantifies the information density in that particular area (Miller and Murphey, 2013).

Geometric features, such as segments, circles, or distances extracted from a map of the environment, are used to define the set of points to consider as candidate viewpoints (Gao et al., 2019).

The Random generation family of approaches includes the Rapidly-exploring Random Trees (RRT) (LaValle, 1998) and other area sampling approaches. A specific advantage of this class of algorithms is that they can deal with high-dimensional and constrained problems.

The Visual features family includes approaches that extract points or areas of interest from camera images. They are used to sample candidate positions that the robot can explore to proceed with the exploration task as in Cimurs et al. (2022) and Yang et al. (2024).

Viewpoint Selection is the second step in an exploration strategy (see the list of papers in Table 9). It consists of identifying among the selected viewpoints the best candidate according to some optimality criteria.

Cost/Utility function is the most used approach to select the best viewpoint. The function can be based on different cost factors like the distance with respect to the candidate positions (Wang et al., 2019c) or the energy required to navigate toward certain areas. The utility is typically measured as the number of frontier cells or the volume of unknown space that can be seen from a viewpoint.

A more specific approach to evaluate the utility of visiting a viewpoint consists in computing the Information gain (Henderson et al., 2020). Here the objective is to select the candidate that maximizes the information acquired by visiting it. Ideally, to evaluate the information gain associated with the candidates, the robot should take into account future (controllable) actions and future (unknown) measurements (Lluvia et al., 2021). In practice, concrete approaches use heuristics such as the expected Shannon’s entropy of the map given a new sensor observation (Bai et al., 2016) or the Cauchy–Schwarz quadratic mutual information (Charrow et al., 2015).

Originally, Potential fields were used as a technique for reactive navigation (Wang et al., 2017). In the context of exploration, potential fields are used, for example, by Yu et al. (2021) to guide the robot through the space by being repulsed from occupied areas and attracted to the frontiers of unexplored areas.

More recently, a technique that has been applied to robotic exploration is Reinforcement Learning (RL), a sub-field of machine learning that enables an agent to learn behaviors through trial and error interactions (Kaelbling et al., 1996). The use of RL methods for exploration allows to be more flexible and robust to noise with respect to geometric methods, as stated by Bigazzi et al. (2022). A common way to use RL for Viewpoint Selection is the identification of an optimal policy to maximize information gain as in Zheng et al. (2022b). The approach presented in Tao et al. (2023) uses deep learning to predict map completion and information gain.

End-to-end approaches use deep learning techniques to learn the mapping between perception and control like using camera image observations to guide a robotic arm for closing a bottle (Levine et al., 2016). In particular, Reinforcement Learning algorithms explore the space of possible control strategies and select the strategy that maximizes a given reward function. End-to-End Approaches accomplish robot exploration tasks as a black box as they do not distinguish between viewpoint generation and selection. Strengths and limitations of reinforcement techniques applied to robotics have been analyzed by Kober et al. (2013).

Similarly, end-to-end approaches for robot explorations feed a learning algorithm with sensor data to compute the control action that guides the robot to perform the next exploration step. We classify end-to-end exploration strategies according to three learning methods (see Figure 4).

Value-based RL methods use a value function to estimate how advantageous is for the robot to perform a control action.

Actor-Critic RL methods split the policy search process into two phases: The actor predicts the actions, and the critic computes the value function.

Imitation learning (Hussein et al., 2017) is an alternative to RL. Instead of optimizing a reward function to select the optimal control policy, the imitation learning algorithm is trained to imitate a provided expert policy.

4.2.3. Multi-robot coordination

Multi-robot coordination is a part of robotic exploration involving the use of multiple robots that collaborate to efficiently explore and navigate the environment. The use of more than one robotic platform is advantageous when the environment is very large or has a complex structure that requires robots with different motion capabilities. In this case, to avoid overlapping, the exploration strategy should also include multi-robot coordination. In the proposed taxonomy, we distinguish between Coordination Strategy and Coordination Technique.

In the primary studies, we have identified two coordination strategies. Distinct areas refers to partitioning the environment in areas that are explored by a single robot, like in Yu et al. (2021). Distinct roles refers to having different robots that execute different tasks, like in the case of a leader robot that collects sensory data, computes a common map to share with the other robots, and plans for all the others as proposed by Griffiths Sánchez et al. (2018).

Coordination techniques refer to the assignment of exploration tasks to each robot. Voronoi Graph decomposes the environment in distinct areas to be explored by individual robots. Potential fields repulse each robot from navigating in areas occupied by other robots. Reinforcement Learning is used by Yang et al. (2024) and Luo et al. (2019) to assign frontier nodes to each robot in a topological map.

Optimization methods are based on a global team manager that assigns different exploration tasks to the robots of the team, in such a way that some performance metric is optimized, such as the total covered path, the duration of the exploration task, the workload balance among the robots, or a more general cost/utility function.

Market-based approaches allow the robots to trade tasks and resources with one another in a decentralized manner in order to maximize the efficiency of the team.

Multi-robot coordination is also concerned with the methods to efficiently share map data among the robots involved in the exploration task in order to limit communication bandwidth. In our classification, we distinguish between Map Sharing and Map Updating.

Map Sharing indicates where the global map is stored and we have identified two approaches. The Centralized approach is used when the robots communicate with a base station that stores the updated global map and coordinates the tasks of the robot’s team, such as planning their paths to the next viewpoint. The base station could be also one of the robots in the team, which has significant computational resources, such as in Qin et al. (2019), where an UGV builds a coarse map of the environment with its onboard sensors and receives fine map data from an UAV. The Decentralized approach is used when all the computation is performed onboard each robot, each of which holds a copy of the updated global map.

Map Updating specifies the strategy to limit the amount of data that each robot has to share to update the global map with new measurements from its onboard sensors. From the analysis of the primary studies, we have identified the following categories.

Each robot communicates specific types of data. Topological data may correspond to frontiers extracted from locally updated occupancy maps, or edges and nodes of a local topological map. Only differences refers to the portion of the local map that has been updated with the most recent sensor measurements. Low resolution means that some data compression technique is used to reduce the size of data to be transmitted by each robot.

Alternatively, each robot communicates only when possible. Only to neighbors indicates that each robot transmits data only to the closest team members because there are communication barriers. Low rate indicates that every local map update is shared with other robots, but this imposes a severe limit on the update rate.

Some approaches combine multiple updating strategies. For example in Hu et al. (2020) and Masaba and Li (2021), robots communicate Topological data and Only to neighbors.

We have found only one approach that combines Centralized map sharing and Only to neighbors communication. This is the case of Williams et al. (2020), where a team of robots explore a subterranean environment with communication barriers. Every robot returns to the base station periodically to upload map data that are merged and used to coordinate the multi-robot team.

5.2.1. Application requirements

The tables in Figure 6 show the most recurrent patterns of application requirements in the selected primary studies.OPEN IN VIEWER

It should be noted that the various types of environment included in the Taxonomy of Figure 3 result from the analysis of the experimental scenarios described in the primary studies. The large majority of the proposed exploration strategies are tested in only one environment, either real or simulated, with only one type of sensor and platform. This does not necessarily mean that the same strategy is not suitable for the exploration of other types of environments with different sensors and platforms. Some authors claim that their approach is generic, but do not provide evidence of it.

Each table corresponds to a type of environment and is structured in three main rows, which correspond to different types of robotic platforms. For the sake of clarity, facet UGV groups both single UGV and multi-UGV systems, facet UAV groups both single UAV and multi-UAV systems, while facet Hybrid refers to hybrid UGV and UAV systems. For the same reason, we have omitted the Underwater environment because we have found a limited number of approaches. Each of the three main rows groups four rows, which correspond to the different types of sensors. Here, Depth Camera includes also ToF Camera and Structured Light cameras. Each table has two columns, which refer to the type of task, that is, Goal-guided and Area coverage. The number in each cell indicates how many primary studies address the specific combination of application requirements.

The majority of approaches have been tested in office-like indoor buildings (65%), where typically research labs are located, with both simple and multi-robot systems of UGVs and UAVs equipped with an entire variety of sensors. In particular, 2D Lidars (54%) are popular because they have an affordable price and good performance in indoor environments. Similarly, depth cameras (31%) combine the advantage of low cost with the possibility of acquiring semantic information about the scene from RGB images. Sonar (11%) has been used in office-like environments mostly for historical reasons.

A number of approaches have been tested in outdoor fields (15%). Indeed, several robotics applications, such as search and rescue, disaster response, agriculture, and monitoring, require the employment of robots in open-air environments.

Subterranean environments are also well represented in our SMS (12%). According to Azpúrua et al. (2023), a variety of sensors can be used for the exploration of subterranean environments. Interestingly, our SMS shows a significant prevalence of 3D Lasers (73%), despite them being the most expensive and energy-demanding type of sensor. This is due to many advantages offered by LiDAR sensors, such as being active sensors with high range, high accuracy, and high density. Other sensors that are commonly used in subterranean environments are 2D Lasers (16%), Stereo Cameras (14%) and Thermal Cameras (8%) and ToF Cameras (5%).

Only a few approaches have been tested in home environments (3%). This is probably due to the difficulties in arranging the typical home furniture in a research laboratory and in conducting experiments in a real home environment due to safety constraints. Nevertheless, service robots for home environments, such as floor cleaning devices, require the ability to build a map autonomously to perform their tasks efficiently without human support. Home environments present challenging characteristics (e.g., moving and movable objects, changing lighting conditions) that demand robust and cost-effective solutions to autonomous robot exploration. The existing gap in the analyzed studies will certainly be filled in the near future, given the development of robotic systems working alongside humans in domestic environments as robot caregivers or collaborative humanoid assistants.

Surprisingly, we have found very few approaches for the autonomous exploration of collapsed buildings (1%). At the same time, we have encountered several approaches for search & rescue tasks that have been tested in office-like or outdoor environments (69%). Certainly, this is due again to the difficulties in conducting experiments in real environments.

Regarding planetary exploration (1%), our SMS reports only a few occurrences since these kinds of studies are a niche in the research community. Moreover, most of the time, the exploration is not completely autonomous since a remote operator usually provides the plan the robot needs to follow on a global map.

A few approaches (7% in total) have been tested in more than one environment, such as Office-like and Outdoor Field (4%), Office-like and Subterranean (2%), Outdoor Field and Industrial Plant (1%) and only 2% of the approaches address the problem of exploring a dynamic environment.

Relatively few approaches enable the robot to perform tasks different than simple area mapping during an exploration mission: Search for Objects (5%), Search and Rescue (4%), and Target Pursuing (3%). This is because, typically, for more complex missions, it is assumed that the map of the environment has been built first.

To conclude the analysis of application requirements, it should be noted that only 20% of approaches are based on multi-robot systems.

5.2.2. Exploration strategies

The tables in the upper part of Figure 7 show the most recurrent patterns among the facets used to classify two-stages exploration strategies documented in the selected primary studies.OPEN IN VIEWER

For each type of map each row reports the number of primary studies for viewpoint generation and viewpoint selection method. We have grouped in a single table Hilbert Maps, Truncated Signed Distance Field, and Gaussian Processes because they are variants of volumetric maps, we have grouped 2D Topological and 3D Topological maps because the latter have been used in a very limited number of approaches, and we have grouped 3D Mesh and Elevation maps because they are specific types of 3D geometric maps.

As described in Section 4.2, we have grouped facets Frontier Detection, Geometric features, Voronoi Graph, and Ergodic trajectories in a single category for VP Generation named Area partitioning as they represent different techniques for partitioning the free space in regions, where promising viewpoints can be found. Similarly, we have grouped facets Random Sampling and Rapidly-exploring Random Trees in a single category named Random generation.

With regards to Area Partitioning techniques for viewpoint generation, the Cost-Utility Function emerges as the predominant approach for viewpoint selection in 2D topological maps (86%), 3D topological maps (83%), 3D geometric maps (62%), 2D occupancy grids (69%), and 3D occupancy grids (54%).

We contend that its prevalence is due to the possibility to easily define cost and utility parameters related to the structure of the environment (e.g., the distance to the centroid of a Voronoi region, the length of an ergodic trajectory, and the size of a frontiers cluster) and to easily balance different cost and utility parameters (e.g., the cost of a viewpoint is high if it is outside a room that has not yet been fully explored). In addition, Area Partitioning techniques have been defined since the foundation phase of the research on robot exploration. Conversely, when viewpoints are generated using Random sampling approaches, techniques based only on the computation of the Information Gain are more commonly adopted.

Interestingly, Visual features viewpoint generation is employed in relatively few works (4%). Only one of them generates Semantic maps (for Underwater environments) (Girdhar and Dudek, 2016) and only one generates 2D Topological maps (for Office-like environments). We argue that other approaches for Office-like environments could be extended to exploit visual features also for the identification of topological elements of the environment and their semantics (rooms, corridors, passages).

As one would reasonably expect, 2D/3D Topological maps are mostly used in combination with Area Partitioning viewpoint generation. Viewpoint selection approaches based on Potential Field (4%) and Reinforcement Learning (7%) have been applied to robotic exploration relatively recently and are primarily associated with 2D Occupancy maps and Area Partitioning.

On the contrary, Semantic maps are equally used in combination with Area Partitioning and Random viewpoint generation.

The tables in the lower part of Figure 7 show the most recurrent patterns among the facets used to classify end-to-end exploration strategies documented in the selected primary studies. Here, the three rows report the number of primary studies for each type of map and for each learning method. Our analysis of the primary studies shows that End-to-end approaches for exploration tasks are not yet as numerous in the literature, especially for non-guided tasks, whose main goal is to cover a whole area. This is coherent with the analysis reported in the survey by Garaffa et al. (2023), where it is shown that End-to-end approaches are mostly used for tasks whose goal is to find or reach a specified target. Additionally, our analysis of the primary studies shows that most approaches are based on traditional types of maps, that is, occupancy grids and topological maps. Unsurprisingly, some End-to-end approaches exploit Semantic maps, in particular for search and rescue tasks, such as in Pham et al. (2018) and Petříček et al. (2019), because RL algorithms can efficiently deal with a large number of states.

5.2.3. Correlations between application requirements and exploration strategies

In this section, we aim to analyze the recurrent correlations between application requirements and exploration strategies.

For this purpose, Figure 8 shows a table with seven main columns, one for each type of map, and four main rows. The first three rows correspond to VP Generation methods of two stages approaches and the fourth row corresponds to end-to-end methods. Thus, main rows and main columns correspond to categories of the exploration strategies taxonomy.OPEN IN VIEWER

Each main column aggregates four internal columns, one for each type of robotic platform, where Hyb stands for hybrid platforms. Here, we do not distinguish between single robots (e.g., UGV) and homogeneous multi-robot systems (e.g., UGV-UGV).

To improve the readability of the tables, we have omitted the columns without entries. Each main row aggregates the types of environments for which at least one exploration strategy was documented in the primary studies. Thus, internal columns and rows correspond to categories of the application requirements taxonomy.

Data in this table show that 2D and 3D occupancy grids have been used to explore every type of environment and with every type of robotic platform. This is probably due to the fact that occupancy is among the first types of maps used in robotics and a significant variety of open source libraries are based on occupancy grids.

Interestingly, 2D/3D Topological maps are mostly (65%) generated while exploring indoor environments with UGV platforms. This result seems reasonable for large structured environments that can be naturally described in terms of rooms, halls, and passages. Indeed, the large majority of these approaches use Area partitioning techniques for viewpoint generation. Beside Office-like environments, several approaches use 2D/3D Topological maps for the exploration of subterranean environments, as they are typically characterized by long underground tunnels as discussed, for example, in O’Brien et al. (2023) for multi-UGV-based exploration or in Yang et al. (2021) for UAV-based exploration.

We found 20 primary studies, where Semantic maps are used for the exploration of almost every type of environment, with the exception of Collapsed and Planetary environments. Most of them refer to Area Mapping (35%) or Terrain Learning (35%) tasks, while the others refer to Search for Objects (20%) and Search&Rescue (10%) tasks.

Viewpoint generation based on Area partitioning is predominant (66%) for every type of environment and every type of platform, Random generation is well represented (27%), while Visual features are seldom exploited (4%).

Interestingly, in the primary studies we did not find any experimental scenarios evaluating random viewpoint generation techniques in Collapsed, and Planetary environments. We do not see any specific reason for this result, which could be considered as a gap that will be filled by future research.

As expected, UAVs, whether single-robot or multi-robot systems, are almost always used for generating 3D maps. Similarly, UGVs are typically used to build 3D maps of open outdoor fields and subterranean environments.

Interestingly some approaches use UGVs to generate 3D maps in office-like environments (5%). In these cases (e.g., Li et al. (2020c)), the onboard sensors mounted on a fixed position do not allow the robot to fully perceive the 3D environment, where objects may be partially visible from a predefined height. For these approaches, the goal is not to build a complete map of the 3D environment but only a 3D map that the robot can use for subsequent navigation tasks.

Surprisingly, only one approach for 3D mapping with an UGV has been tested in home environments (Zurbrugg et al., 2022). This is probably due to the current availability of simple commercial devices like floor cleaners, which only need 2D maps for navigation. If multi-legged robots (e.g., robot pets), mobile manipulators, and humanoid robots will become affordable commodities, they certainly will need autonomous exploration capabilities for building accurate 3D maps for a variety of navigation and manipulation tasks.

The random generation of viewpoints has been used in very few approaches to building topological maps because structured approaches (such as Area partitioning) offer a better approach regarding efficiency and map completeness.

End-to-end approaches have been mostly developed for UGV platforms in easily accessible environments, that is, Office-like and Home environments, where the experimental tests can be conveniently performed with real robots. One possible reason is that end-to-end methods tested in the real world typically use computationally heavy CNNs, which require adequate onboard computing resources.

5.2.4. Multi-robot coordination

The table in Figure 9 shows the most recurrent patterns related to multi-robot coordination in exploration tasks. The table has six main columns corresponding to the map types defined in the exploration strategies taxonomy (see Figure 4). Semantic maps are not included in the table because we have not found any multi-robot exploration approach that produces a semantic map. For each map type, two distinct columns correspond to different coordination strategies, that is, Disinct Areas and Disinct Roles, which are further detailed in two columns indicating the type of platform, either homogeneous or hybrid.OPEN IN VIEWER

The whole table is structured into three main rows, corresponding to the Coordination technique, the Map sharing, and the Map updating methods.

Data in this table show that the coordination strategies employed for multi-robot exploration are mainly obtained with robots allocated in distinct areas of the environment (83%) rather than with distinct roles (17%). As expected, this prevalence is associated with the prevalence of homogeneous multi-robot-systems (87%) over hybrid multi-robot-systems (13%). Nevertheless, there are coordination approaches that assign distinct roles to robots with the same type of platform, as in the work by Mahdoui et al. (2020), where a specific robot of the team plays the role of leader and is in charge of making cooperative decisions. Similarly, there are approaches that assign the exploration of different areas to robots with different platforms, as in the case of subterranean environments that require hybrid multi-robot systems with heterogeneous capabilities (Best et al., 2023).

On top of this consideration, it can be seen that the most employed coordination strategies, regardless of the type of map used, optimize the covered path (40%) and some cost-utility function 25%). It is worth highlighting that a significant number of studies (13%) employed market-based techniques to coordinate the robots during the exploration.

With regards to Map sharing there is a slight prevalence of Decentralized (59%) over Centralized (41%) methods. This prevalence is higher for 2D Occupancy grids and Topological. 2d Occupancy maps are the most commonly used map type when robots are tasked with exploring different areas and they are mostly shared uncompressed at low communication rate or only to nearby robots.

It seems reasonable that Neural implicit maps are not commonly used for the exploration by multi-robot systems because RL algorithms require adequate onboard computing resources. Notably, only the approach proposed by Yang et al. (2024) builds a topological variant of Neural implicit maps, which are stored on a central server.

Similarly, it seems reasonable that semantic maps are not commonly used in multi-robot exploration because they typically store a large amount of information about the environment characteristics, which would require a high communication bandwidth among robots.

Interestingly, there is a significant number of approaches that share topological data used only for coordination, which are extracted from locally updated 2D / 3D Occupancy grids.