A categorization of simultaneous localization and mapping knowledge for mobile robots

Autonomous robots are playing important roles in academic, technological, and scientific activities. Thus, their behavior is getting more complex. The main tasks of autonomous robots include mapping an environment and localize themselves. These tasks comprise the Simultaneous Localization and Mapping (SLAM) problem. Representation of the SLAM knowledge (e.g., robot characteristics, environment information, mapping and location information), with a standard and well-defined model, provides the base to develop efficient and interoperable solutions. However, as far as we know, there is not a common classification of such knowledge. Many existing works based on Semantic Web, have formulated ontologies to model information related to only some SLAM aspects, without a standard arrangement. In this paper, we propose a categorization of the knowledge managed in SLAM, based on existing ontologies and SLAM principles. We also classify recent and popular ontologies according to our proposed categories and highlight the lessons to learn from existing solutions.


INTRODUCTION
Nowadays, autonomous robots are playing important roles in every day life activities, as well as for academic, technological, and scientific applications [9,15]. The main tasks of autonomous robots include mapping an environment and localize themselves. These tasks conform the Simultaneous Localization and Mapping (SLAM) problem. Thus, SLAM handles the need of constructing a map of a unknown environment and simultaneously determining the location of the robot within the built map.
In general, building maps of unknown environments is based on information captured by a range of sensors, such as lasers or sonars, while robot location needs more information coming from other types of devices (e.g., GPS). Due to the evolution of mobile and sensor technologies, the behaviors and tasks that robots can perform are becoming more complex, including the possibility of collaboration among multi-robots and among humans and robots. Thus, this trend also demands the use of more complex knowledge.
Even though, SLAM is an area that has been the focus of many works, thus reaching a very good level of maturity [17], there is still a lack of standardization to represent such information and the knowledge needed to propose efficient and interoperable solutions. In this sense, it is evident the necessity of a well-defined and standard model for modeling the knowledge managed by SLAM algorithms. Semantic Web seems to be a clear solution, from which we can take its organizational and relational capacity. However, as far as we know, there is not a common classification of such knowledge. In fact, many existing works have formulated ontologies to model information related to only some SLAM aspects [4,13,16,22,34], without a standard and common arrangement.
To overcome this limitation, in this paper, we propose a categorization of the knowledge managed in the SLAM problem, based on existing ontologies and SLAM principles, in particularly, for mobile robots. We also review, analyze, and classify recent and popular SLAM ontologies according to our proposed categorization. Based on this analysis, we explore new data fields to obtain criteria for enrich the existing ontologies and highlight new possibilities in terms of performance, efficiency, and design better solutions for SLAM applications.
The remainder of this paper is structured as follows. Section 2 describes some studies related to our work. Section 3 presents the most important concepts of SLAM. Section 4 presents our proposed SLAM knowledge categorization. Section 5 develops a SLAM ontologies classification based on the proposed categories. Finally, in Section 6, general conclusions of this work are presented.

RELATED WORKS
Some works have focused on the standardization of robotic terms, mostly based on taxonomies and ontologies, and discuss their significant impact on mobile robotics. However, they have covered only partial topics for the SLAM problem. As far as we know, there is not a previous work dealing with the knowledge need for SLAM problem as a whole (robot characterization, mapping, localization, and specif domain information).
The studies presented in [13,34], highlight the importance of standards and ontologies to support general mobile robotics development and present reviews of robotic ontology projects for those moments (2013 and 2016, respectively). However, in these studies the focus is how representing the knowledge mainly related to service robots (i.e., robots characteristics and capabilities). Thus, there is a lack of other important knowledge related to SLAM (e.g., environment, location, mapping).
In [4], authors classify ontologies in three classes: (i) for general robotics applications; (ii) for autonomous robots; and (iii) for autonomous vehicles. The second class is the most related to the SLAM problem, considering knowledge related to robot environment, description and reasoning related to robot actions and tasks, and capabilities of robots. In [22], the aim is to report how learning techniques (deep learning, probabilistic modelling, machine learning, and semantic graphical structures) in service robotics deal with knowledge representation. According to authors, ontologies play the most important role on high level knowledge representation of task for the process of robot learning, but information related to the the description of the robot environment is neglected.
A study focused on reviewing ontologies related to the mapping aspect of robots, which are able to provide an abstraction of space, i.e., a qualitative description of robot environments to augment the robot task-planning and navigation capabilities, is presented in [16].It describes the trends on semantic map building. Authors classify the reviewed works according to the capacity of modelling indoor single scenes, indoor large-scale scenes, outdoor single scenes, or outdoor large-scale scenes. However, this work does not deal with information related to the robot aspects.
Even thought, we have common classifications with some of these works, about the knowledge needed for SLAM problems, they deal with general ontologies for generic service robotics. In contrast, we consider all aspects related to the information managed in SLAM applications, by proposing a classification of this knowledge. We also present the most up-to-date and the most focused review on ontologies related to SLAM, that could be useful for researchers and developers of mobile robots in the field of SLAM.
A quite good and very recent survey about SLAM is presented in [29]. This study presents a good state-of-art in SLAM solutions, the most used techniques in SLAM systems implementation, and discusses the insights of existing methods. It includes a review of semantic SLAM approaches that concerns the semantic information into the SLAM process. However, ontologies are not considered as an approach to incorporate semantic to the SLAM information. Nevertheless, we believe that this study complements our work in this paper. That study presents the general methods to resolve the SLAM problem and we provide the view of semantic information based on ontologies, to enhance its representation by providing robust performance, task driven perception, resource awareness, and high-level understanding.

SLAM PRINCIPLES: PRELIMINARIES
SLAM is a well-known problem by which a mobile robot must construct a map of a specific environment and simultaneously identify its own position within this map [11].
To build a map, a SLAM solution is mostly based on the identification of some representative objects in the environment, called landmarks. Landmarks are static objects, that can be used to identify places or zones in the environment. An efficient and consistent solution for SLAM requires to combine the robot pose and every landmark's position, according time information, in a unique state. The state must be updated on every robot observation, considering that it includes certain uncertainty [11]. There exist three kinds of maps that can be constructed, either in 2D and 3D, when solving the SLAM problem: (i) join pose-feature maps; (ii) pose-only maps; and (iii) feature-only maps. A pose-feature map is a map which consists of the landmarks and robots positions. A pose-only map is a map consisting of relative positions among the robot poses. A feature-only map considers all landmarks features [35]. Building an accurate map, along with performing precise localization of robots is a non easy problem, that demands a variety of information [2,11,35]. The literature coincides that the information that should be considered is mainly related to: (i) the map itself; (ii) the robot pose; (iii) semantic and features of the workspace; (iv) changes of poses and landmarks features on time; (v) data representing uncertainty; and (vi) some information related to the kind of solution is implemented [2,11].
A representation of the map itself is, of course, the main information that must be stored. It includes geographical information about all the environment with landmarks correlation. It is important to storage features representing the shape and position of landmarks. Also, it is important to consider that this information comes from an estimation and this estimation can change on time. Finally, it is important to consider and storage specific domain information of the environment, either if it is obtained for each landmark or for zones in the environment.
Robot pose is another important information to count on, because the problem is about localization while mapping an environment. It is important to include kinematic information of the robot to understand its pose and the space it holds. Also, for understanding how the map is estimated, and possibly improving that estimation, it is important to storage sensory information and the robot trajectory. Robot trajectory is relevant for some kinds of SLAM problems, like active SLAM, which considers the capability of mobile robots to generate on-time trajectories to maximize the accuracy of the generated map and the localization of robots [7].
Beliefs about the map and positions of objects will change every time; then, it is important to store these changes including the information of moving objects. These beliefs changing will bring uncertainty that must be represented, quantified, and stored at each time. Stored uncertainty will bring the level of confidence associated to the estimation of robot poses and mapping [27]. If the solution is a large scale one, it could include sub-mapping and delayed mapping, then it is important to store correlated sub-maps and relative correlations of positions, shapes, and poses.
According to these SLAM principles, some works have proposed to partially model the information managed in SLAM applications, by using ontologies. This way of modeling the knowledge in SLAM is inefficient because it is not standardized and is incomplete. In this sense, it is important to address this limitation by proposing a standard categorization of the knowledge managed in the SLAM problem (Trying to include the most categories of knowledge as possible). So, By analyzing and summarizing the current state of art in SLAM, we propose a categorization of such knowledge, considering all aspects related to SLAM solutions.

SLAM KNOWLEDGE CATEGORIZATION: OUR PROPOSAL
It is obvious that in a SLAM scenario, there exist different types of information and knowledge that are managed, considering the description of the capabilities, characteristics, behaviours of robots, as well as environmental and specif domain information.
In order to correctly categorize the knowledge in SLAM, it is important to consider that: (i) SLAM is a problem that must be solved by autonomous robots; (ii) a SLAM solution is a continuous solution; thus, the problem is not completely solved after some time; that means that while the robot is working, it must map the environment and locate itself; (iii) since a solution to SLAM is continuous in time, then robots will always have some "uncertainty" about the correctness of the mapping process and also about its location; (iv) the mapping process and the robot's location estimation depend on the correctness about believes of shapes and location of landmarks; (v) the location of a robot depends also of its physical structure, defined by its kinematics; (vi) it is important to consider that the main goal of autonomous robots is to act in "real world", that means, to act in dynamic environments with possibly moving objects (passive or active) that must affect the process of mapping and location; (vii) in order to improve the robot capabilities of selflocation while mapping an environment, it is important to manage information and knowledge about physical and semantic characteristics of landmarks in the environment; and (viii) information and knowledge about the environment depend on the dimension of the perception of the robot (2D or 3D) and also in the specific domain of application of the robotic solution is implemented.
Hence, based on these considerations, our proposed classification considers the following fields of SLAM knowledge: (1) robot information; (2) environment mapping; (3) timely information; and (4) workspace. We detail each category as follows.
(1) Robot Information: It is related to the information that conceptualize the main characteristics of the robot, physical and structural capacities. It mainly considers: (a) Robot kinematic information: It is related to robots' degrees of freedom, degree of mobility, among others. This information is important for modeling the actions that the robot can perform, that in turns allows to consider the navigation of the robot in the space it is discovering. In order to demonstrate the suitability and completeness of our categorization, we perform a comparative analysis of existing ontologies and evaluate which information is modeled on each one. Additionally, we report: (i) the Application Scope in which the SLAM problem was applied (e.g., service, rescue, automatic search), to refer to the context of applicability of the ontology; and (ii) the Origin Ontology (Based on), if the ontology has been based on an older ontology or on a more general level, such as an Upper ontology. In the next section, we present such comparative analysis.

SLAM ONTOLOGIES CLASSIFICATION: A REVIEW
In order to find, select, and analyze the most popular and recent ontologies related to SLAM, we have followed a systematic review consisting of three main steps: (i) search of works dealing with ontologies related to SLAM; (ii) selection of relevant articles; and (iii) elaboration of a comparative analysis based on our proposed SLAM knowledge classification. For the first step, the search was done on the search engine of Google Scholar, which provides links to scientific repositories such as IEEE Xplore, ACM, and Springer. The search started with the tags "SLAM ontology" and "CML ontology" (since SLAM is also called Concurrent Mapping and Localization), but no jobs were found that fully coincide with the search. Thus, we extended the search field to "mapping ontology" and "semantic mapping". With these tags more jobs were found. We obtained around 40 articles.
In the second step, we selected the most relevant articles related to specific SLAM information. From the more than 40 scientific papers obtained in the first step, some of them do not propose ontologies, others propose ontologies but not in the specific area of SLAM. We selected works from 2010 and some older ones that have been widely cited or represent the basis of more recent projects. The final result was 19 relevant papers.
In the third step, we analyzed and compared these 19 papers, according our proposed categorization of the information managed of SLAM applications. We detailed how those selected ontologies model SLAM specific information, in order to know which are the most developed fields and which are still missing.

Robot Information
Before a robot can describe and know its surroundings, it needs to know what it can do, which are its physical, structural, and functional capabilities. In this section, we describe ontologies that are mainly focused on those aspects, however some of them also take into account some aspects related to the environment, mapping, and specific domain knowledge.

Robot kinematic information:
Even though the RoboEarth ontology [26] main focus is to model concepts and relationships among objects and maps, it is the best one we found that provides a good kinematic robot model and robot's motion capabilities model (e.g., it can represent the arm motor capacity). Robot Ontology [28], models a neutral knowledge about robots and their capacity to help in the field of robot search and rescue systems. This ontology model structural characteristics, functional capabilities, and operational considerations of robots. All these data are mainly captured in the definition of the robot itself, including: size, weight, power source, sensors, and processors. With this information, it is possible to conceptualize locomotion, sensor, operational capabilities, and also degree of autonomy.
The work presented by Burroughes and Gao in [5], proposes an ontology which dedicates an entire module to represent the needed information necessary for the self-reconfiguration of a planetary rover. Clearly, this is a very broad and challenging domain. To reduce the intrinsic complexity, the ontology is organized in modules. The Upper ontology is divided into sub-modules. Another point of view of kinetic information is to model actions of robots as is the case of: (i) OASys ontology [1,23]; (ii) the ontology used in OUR-K (Ontology-based Unified Robot Knowledge) [19]; and (iii) KNOW ROB ontology [31]. For these three ontologies, in order to represent these complex actions, it is necessary to have a previous knowledge of the locomotive capacities of the robot.
Another group of ontologies have not focused on describing the robots' locomotive capabilities but they have classes and entities to describe robot parts. This is the case of CORA [24], POS [6], and the work of Fortes-Rey [25], which have been inspired by the general concepts of SUMO [12] and have a RobotPart entity in common.

Robot sensory information:
In SLAM applications, sensors from which gather information can be present in the environment or in the robot. KNOW ROB ontology [31] models information received by sensors placed in the environment. A remarkable feature of this ontology is the management of uncertainty, that could be caused by hallucinated objects detection,limited observability and sensor noise. ROSPlan [8] manages both sources of information (i.e., sensors in the environment and in the robot) to locate valve panels and valves. These objects are important for the case study of the work. Most considered ontologies model information received from the sensors in the robot. Some of these ontologies have a whole ontology only for sensors, as is the case of SUMO [12] and the ontology proposed by Burroughes and Gao [5].
Other ontologies model sensory information, even it is not their main objective, such as: (i) RoboEarth ontology [26], (ii) the ontology of SEMAP, a framework for semantic maps representation in spatial databases, proposed by Deeken et al., in [10]; (iii) the ontology proposed by Wu et al., in [33]; (iv) the ontology used in OMRKF (Ontology-based Multi-layered Robot Knowledge Framework) [30]; (v) based on OMRKF, the knowledge framework OUR-K [19] is presented to be used for service robots, (vi) OASys ontology, that treats sensors as a type of device; it is developed for Unmanned Aerial Vehicles (UAV) and Unmanned Ground Vehicles (UGV); and (vii) Robot Ontology, where the structural characteristics are captured when the robot is defined; one of this characteristics is related to sensors (e.g., camera, Temperature Sensor, GPS, SONAR, Audio).

Robot pose information:
Robot position can be absolute or relative. For the first case, a robot is positioned at a global spacial coordinate (e.g., (x, y) if its workspace is R 2 ), while in the second case, the robot is positioned considering the position of a landmark in the environment (e.g., the robot is behind the "wall", where behind means a conical region centred on the "wall" and pointing backward). In the absolute case, works like: (i) Burroughes and Gao's proposal [5], that with its Topological Map Ontology and Simple Map Ontology can represent an absolute position; (ii) SUMO [12], which we assume it models absolute positions, in a coordinate system, because in one application in which it is tested, it is possible to model car positions in a New York map; and (iii) the ontology presented by Wu et al., in [33], which uses a Bayes algorithm for a reliable position of the robot in spatial semantic hybrid map building; it is a more accurate way of modeling robot position. On the other hand, works able to represent relative positions are: (i) RoboEarth [26]; (ii) the work of Fortes-Rey [25], that can represent both absolute and relative positions; (iii) POS Ontology [6], which complements Core Ontology and expands SUMO, by specifying the main concepts and their relationships, and underlying the notions of pose, orientation, and position; additionally, POS allows the description of postures, orientations, and positions, in a coordinated system; in this way it is possible to model quantitatively and qualitatively orientations and positions; and (iv) the ontology proposed by Deeken et al., in [10] for semantic maps in spatial databases, whose mapping approach consists on obtaining absolute geometric data from the environment to model objects and their relative spatial relationships.
There are proposals that do not generate a metric map as a result of their SLAM stage, in contrast they generate a graph where each node represents a possible place or position where the robot can be. The Knowledge Base ontology of ROSPlan framework [8], uses the structural similarities among many robotics planning domains. For example, authors represent the areas, where an Autonomous Underwater Vehicle (AUV) can move, showing with points its possible locations (waypoints). In a similar way of [8], KNOWROB [31] models robot pose in a list of places where the robot could be, using the RobotPlace entity; and the ontology proposed by Hotz et al., in [14], which combines two spatial reasoning calculi, RCC-8 and CDC, with ontological representations of maps for service robots for a restaurant. There are also ontologies which combine different maps like OUR-K ontology [19], that combines semantic, topological, and metrics maps to describe Spaces; and the ontology for multi-layered conceptual maps proposed by Martinez et al., in [20] and the one proposed by Li et al., in [18], which represent metric, navigation, topological, and conceptual maps.

Robot trajectory information:
Considering that a trajectory is a sequence of positions in a given time, an ontology has to model time to properly model trajectories. Two ontologies comply with this temporal characteristic, the one proposed by Burroughes and Gao [5] and the proposal of Fortes-Rey [25]. The first one has a module called Proccess Ontology related to Temporal Ontology. Instead, Fortes-Rey proposes a relationship, called posAtTime, that allows to relate a robot to a position and timing. Then, it is possible to describe the trajectory of a robot.

Robot position uncertainty:
Only two ontologies were found that conceive uncertainty. KNOW ROB ontology [31]. This ontology considers the actions of robots are unreliable and inaccurate. Also the ontology for Spatial Semantic Hybrid map building proposed by Wu et al., in [33], which is able to represent the most probable position where the robot can be.

5.1.6
Discussion: Almost all ontologies model basic robot information, such as kinematic, sensory, and pose. This kind of information is most focused on the result of SLAM solution: maps and robot's locations. However, trajectory and position uncertainty information, which are focused on "the process to obtain" maps and locations, are not very considered.

Environment mapping
SLAM deals with the ability of robots to localize themselves in a map/plan and construct the map (outdoor use) or the floor plan (indoor use). Thus, geographical information, as well as information related to landmarks present in the explored space, are needed to keep. We describe ontologies able to model such information.

Geographical information:
In this category of information, RoboEarth ontology [26] is the most outstanding. This is reflected in its ability to model the location of objects with respect to the robot, the geometry of the scene, as well as the relationships and concepts between objects and maps.
There are other ontologies that model geographic information in a general way. This is the case of: (i) Space Ontology [3], which is a spatial knowledge representation complemented with a reasoning system, able to model and manage the space (e.g., hierarchical organizations, spatial entities); (ii) the ontology proposed by Burroughes and Gao [5] has two modules related to geographic information: the Simple Map Ontology and the Topological Map Ontology; and (iii) the ontology proposed by Hotz et al., [14], which describes an environment as a topological graph and separates overlapped and reachable rooms.
Instead of using topology maps, other form to describe an environment is to jointly use metric maps, navigation maps, topological maps, and semantic maps. This is the case of: (i) OUR-K ontology [19], which has a Knowledge Class, specifically to handle spatial notions as metric and topological map; (ii) the ontology presented by Li et al., in [18], which shows how the interaction with an intelligent wheelchair is done by combining multi-layered maps; (iii) the proposal of Martinez et al., in [20], which defines topological areas, defined as an ontological instance of the type Area, in the conceptual map of the ontology (iv) OMRKF ontology, in which Rooms are defined as Spaces; (v) the ontology proposed by Deeken et al., in [10], which models rooms with objects associated with them; (vi) CORA [24], POS [6], and the work of Fortes-Rey [25], which have been inspired by the general concepts of SUMO [12]; they share concepts such as Region and Environment.
Finally, there are ontologies that do not describe the environment as a space where there are objects, but build an environment from the objects. This is the case of the work of Wang and Chen [32] and the one proposed by Wu et al., [33].

Landmark basic information:
When we refer to modeling the basic information of landmarks, we have considered two criteria: (i) the capability of modeling an object other than the robot on the map; and (ii) the capability of modeling the position of this object with respect to the map. Almost all ontologies have defined the entity Object or Artifact to describe landmarks. The work presented by Burroughes and Gao [5] has a whole ontology dedicated to modeling Objects. CORA [24], POS [6], and the work of Fortes-Rey [25], which are extensions of SUMO [12]. RoboEarth ontologie [26], use the Object entity, which is a specialization of the concept Entity, that can be either Abstract or Physical. The ontology proposed by Hotz et al., [14] uses the TBox concept to model objects in the environment (such as cup, plate, table, room). The proposal of Martinez et al., in [20] defines a conceptual map that is the link between the communication system used for the dialogue between the robot and the human when they refer to representations of spatial entities, such as instances of Objects or Rooms, and low-level maps. In a similar way, the proposal of Deeken et al., in [10] has the Object-Description entity,which defines a generalized model of the spatial characteristics for each class of object.
The ontology proposed by Li et al., in [18], describes its environment also with the help of a conjunction of metrical, topological, and semantics maps, just as Martinez's work makes use of relationships has-a. Thus, it is possible to have relations such as Building has-a Floor, Floor has-a Room, and Room has objects like a Desk or a Book. This ontology also allows the modeling of relationships between objects in a Room, such as Book on-a Desk.
RoboEarth ontology [26] and OUR-K ontology [19] can model compound and simple objects, where each object belongs to a position node, associated to an area in a relative (not absolute) way. The method of representing indoor environment semantic maps for mobile robots proposed by Wang and Chen [32], is totally different from its predecessors. For this work, the semantics of an environment does not longer begin with identifying and connecting Spatial Regions; instead, it defines a Region based on the objects that compose it. For example, an office can be defined by the presence of a chair, a desk, a room, walls, or other things.
KNOW ROB [31] and OMRKF [30] ontologies model the absolute positions of objects, i.e., with coordinates (x,y,z). The ontology proposed by Wu et al., [33]models the name, size, function, color, shape, and other relevant data of features of an object, by using a QR code. In the ROSPlan ontology [8], with the data of the sensors captured while the plan is executed and after being collated in the ontology, it is possible to define the resources of the type objects and the relations between them.

Landmark shape information:
This information about landmarks, is important not only in the process but in the final result. Thus, when considering shape information is possible to obtain a more realistic an accurate map of the world. From the studied ontologies, only one explicitly models the shape of landmarks, that is the ontology proposed by Wu et al., [33]. It includes the shape of the landmark, as well as the size, the color, among other characteristics.
Other ontologies model partially this informations, such as: (i) RoboEarth ontology [26], which represents a set of characteristics of the surfaces of the object, with a multi-view geometry; (ii) OUR-K ontology [19], it does not specifically analyze the landmark form, however it offers the possibility of decomposing a landmark in other simpler ones; for example a cup is composed of a body and a handle ; and (iii) the one proposed by Wang and Chen [32], which counts on relations part-of and has-a to describe more complex landmarks; thus, even though the landmark shape is not described geometrically, it does so structurally.

Landmark position uncertainty:
The only ontology that can fully model the uncertainty of landmark positions is the one proposed by Wu et al. [33]. To model this uncertainty, this proposal uses the hidden Markov model and a probabilistic approach based on the Bayes algorithm. RoboEarth ontology [26] also has an approximation to the probability of positions, because based on where the landmarks are located, the knowledge base deduces possible locations, where the objects could be.

Discussion:
Most of the ontologies consider to model geographic information and basic information of the landmarks. Just one of them models information related to the shape of landmarks, while only few of them, partially represent this knowledge. Furthermore, similar to the previous category the number of ontologies that model the uncertainty of landmark positions, is very low.

Timely Information
SLAM is a problem solved by mobile robots and can consider dynamic environment. Thus, not only static positions should be modeled. It is important to model the temporary factor that affects both the environment and the robot.

Time information of robots and objects:
A diachronic ontology can represent state changes of its concepts through the time. SUMO [12] has support for indexing facts over time. In fact, SUMO represents time using TimeMeasures classes, that represent positions or intervals in the universal time-line. In a similar way, the ontology proposed by Fortes-Rey [25], includes the time component in points and regions position measurements. On its side, the ontology proposed by Burroughes and Gao [5] has a complete module to model temporal information (Temporal Ontology). OMRKF [30] and OUR-K [19] ontologies, present a model based on levels, where one level is Context and the temporal context is considered.

Mobile objects:
We can say that the proposal of Burroughes and Gao [5] allows to recognize mobile objects, if we analyze the relationships among the ontologies of Objects and Single Map, with the Temporal Ontology. Also the work of Fortes-Rey [25] allows to differentiate objects that move from those that do not, since the ontology can represent an object placed in a position at a given time. Ontologies such as RoboEarth [26] and OMRKF [30] recognize mobile objects from databases, according to their visual characteristics, instead of their temporal characteristics. For example, a bicycle is considered a mobile object because of its shape but not because of its movement.

Workspace
General characteristics of the work space where the SLAM solution is applied are mainly related to the dimensionality to represent maps and specific semantic knowledge of the domain.

Dimensions of mapping and localization: Some ontologies
show that they can model knowledge related to mapping and location in two dimensions such as the work of Fortes-Rey [25], POS [6], SUMO [12], and OUR-K [19].
As expected, more recent ontologies are already able to model information in 3D. The ontology proposed by Burroughes and Gao [5] allows to receive and model information from an external 3D mapping service. The one proposed by Wang and Chen in [32], defines relationships between objects like,back, right back, right front, left,left back, left front, below, and above, giving the idea of a cube (3D) as a scenario around the object. The same case is presented in OMRKF [30], in the ontology proposed by Li et al., in [18], KNOW ROB ontology [31], and the ontology proposed by Wu et al., in [33].
Others ontologies such as RobotEarth [26] and the one described by Deeken et al., in [10] model 2D and 3D SLAM.

Specific Domain Information:
Once the modeling of information related to the two main SLAM tasks (i.e., mapping and localization), is solved, the next step is regarding the information to detect an object and identify it. It does not only matter identifying objects that exist, but objects that belong to a specific domain. This is the case for the ontology presented by Hotz et al., in [14] and KNOW ROB ontology [31] that have entities only for restaurants. KNOW ROB was also applied in robotic housework [21].
Finally, specific domains less specialized are the indoor environments modeled in: (i) the work proposed by Martinez et al., [20], in which typical indoor environments, such as kitchen, living room, office, and laboratory can be modeled; including also the objects that can be found in them (e.g., blackboards, desks, armchairs, fridges; (ii) OMRKF [30], which is also able to model a kitchen and a living room, including objects such as cups, tables, chairs; (iii) the ontology proposed by Li et al., in [18], which instead models an academic environment, where laboratories, offices, corridors, and computer rooms can be modeled, including objects such as computers and desks; and (iv) RoboEarth [26] provides a sub-database of relevant object models that will be needed to fulfill the target task. Thus, its semantic reasoning enhances recognition by reducing the false positive rate and computation. Table 1 summarize all analyzed ontologies. The ontologies are marked with a cross if they conceptualize the subcategory of the associated column. We identify that some SLAM ontologies do not cover only one category, normally they cover more than one. However, most ontologies consider the first two categories, considering Robot Information and Environment Mapping, while Timely Information and Workspace appear as complement with the others.

General Discussion
SLAM problem is continuous in time; that means that it is desired that an autonomous robot must be solving it all the time, since it works in a real world and it can find new places at any time. We have parsed the importance of representing the timely information into SLAM ontologies, to improve the process of solving SLAM. We consider that the integration of environment mapping information, robot information, and timely information with the positioning uncertainty in robots and landmarks, represent an option to optimize and improve the precision of the results of SLAM solutions.
From Table1 we can note that there is no ontology that models all categories completely. For this reason, as a future work, it is planned to gather the acquired knowledge to model an ontology that satisfies all categories, in other words, that models the entire SLAM process. On the other hand, there are very few ontologies that model important aspects of SLAM, such as uncertainty (Columns 5.1.5 and 5.2.4) and temporality (Column 5.3.1 and 5.3.2).
It is important to develop ontologies for SLAM because more detailed information can take developers to include high level reasoning in autonomous robots in decisions including details of the environment. Additionally, we consider relevant to know the information relative to: (i) origin ontology, to know if the proposal has a predecessor; in this regard we can note that more than a half ontologies are based on an older ontology such as SUO KIF, KAON, or RCC-8, but also we can find ontologies that have been defined from the scratch as Robot Ontology and OASys; and (ii) the application scope, since there are ontologies for several knowledge areas, not limited to Service or Search and Rescue robots; we find other areas such as AUV and planetary robots.
We finally point out the fact that each aspect considered in our proposed categorization is addressed for at least one of the revised ontologies. This indicates that we have took into consideration the most important knowledge related to the SLAM problem. Also, our knowledge categorization allows to evaluate the completeness of SLAM ontologies and to identify lacks and challenges that can boost future research in this area.

CONCLUSIONS
In this paper, we propose four categories for the knowledge managed in SLAM algorithms: Robot Information, Environment Mapping, Timely Information, and Workspace information. Based on this categorization, we analyze several SLAM ontologies and point out that there is not an ontology fully covering the whole SLAM problem. Most ontologies are focused on the results of SLAM (i.e., obtain a map and localize a robot in that map), but not in the process of obtaining those results. The last can be concluded since, from 19 evaluated ontologies, only two include information about robot trajectory and only two include information about robot position uncertainty, but none of them include both characteristics. Almost the same occurs with environment mapping information, here only four ontologies include information about the shape of landmarks and only two include information about landmarks position uncertainty. In this case, those that include landmark position uncertainty and shape information, do not include robot trajectory and robot position uncertainty. The same occurs with the inclusion of time information and mobile objects. Thus, it is important to note that there are not ontologies that can be considered complete solutions to the SLAM problem. As future work, we intend to propose a more complete ontology capable of representing all levels of SLAM information, in order to obtain better SLAM solutions. We plan to extend and integrate the ontologies already existing in this domain. The effort of defining a new general ontology to represent the SLAM knowledge, also represent a strategy to validate our proposed categorization of such knowledge. We will be able to demonstrate the ontology suitability and also measure its performance.