Ontologies have gained popularity in robotics with the growing complexity of actions that systems are expected to perform. A well-defined standard for knowledge representation is recognized as a tool to facilitate human–robot collaboration in challenging tasks (Fiorini et al., 2017). The IEEE Standard Association of Robotics and Automation Society (RAS) created the Ontologies for Robotics and Automation (ORA) working group to address this need.

They first published the Core Ontology for Robotics and Automation (CORA) (Prestes et al., 2013). This standard specifies the most general concepts, relations, and axioms for the robotics and automation domain. CORA is based on SUMO and defines what a robot is and how it relates to other concepts. For this, it defines four main entities: robot part, robot, complex robot, and robotic system. CORA is an upper-level ontology currently extended in the IEEE Standard 1872-2015 (IEEE SA, 2015) with other subontologies, such as CORAX, RPARTS, and POS.

CORAX is a subontology created to bridge the gap between SUMO and CORA. It included high-level concepts that the authors claimed to not be explicitly defined in SUMO and particularized in CORA, in particular those associated with design, interaction, and environment. RPARTS provides notions related to specific kinds of robot parts and the roles they can perform, such as grippers, sensors, or actuators. POS presents general concepts associated with spatial knowledge, such as position and orientation, represented as points, regions, and coordinate systems.

However, CORA and its extensions are intended to cover a broad community, so their definitions of ambiguous terms are based solely on necessary conditions and do not specify sufficient conditions (Fiorini et al., 2017). For this reason, concepts in CORA must be specialized according to the needs of specific subdomains or robotics applications.

CORA, like most of the other application ontologies considered here, is defined in a language of very limited expressiveness, mostly expressible in OWL-Lite, and is therefore limited to simple classification queries. Although it is based on upper-level terms from SUMO, it recreated many terms that could have been used directly from SUMO. Moreover, given its choice of representation language, it did not use the first- and higher-order logic formulas from SUMO, limiting its reuse to only the taxonomy.

The IEEE ORA group created the Robot Task Representation subgroup to produce a middle-level ontology with a comprehensive decomposition of tasks, from goal to subgoals, that enables humans or robots to accomplish their expected outcomes at a specific instance in time. It includes a definition of tasks and their properties and terms related to the performance capabilities required to perform them. Moreover, it covers a catalog of tasks demanded by the community, especially in industrial processes (Balakirsky et al., 2017).

This working group also created three additional subgroups for more specific domain knowledge: Autonomous Robots Ontology, Industrial Ontology, and Medical Robot Ontology. Only the first of these has been active. The Autonomous Robot subgroup (AUR) extends CORA and its associated ontologies for the domain of autonomous robots, including, but not limited to, aerial, ground, surface, underwater, and space robots.

They developed the IEEE Standard for Autonomous Robotics Ontology (IEEE SA, 2022) with an unambiguous identification of the basic hardware and software components necessary to provide a robot or a group of robots with autonomy. It was conceived to serve different purposes, such as to describe the design patterns of Autonomous Robotics (AuR) systems, to represent AuR system architectures in a unified way, or as a guideline to build autonomous systems consisting of robots operating in various environments.

In addition to the developments of the IEEE ORA working group, there are other relevant domain ontologies, such as OASys and the Socio-physical Model of Activities (SOMA). The Ontology for Autonomous Systems (OASys) (Bermejo-Alonso et al., 2010) captures and exploits concepts to support the description of any autonomous system with an emphasis on the associated engineering processes. It provides two levels of abstraction systems in general and autonomous systems in particular. This ontology connects concepts such as architecture, components, goals, and functions with the engineering processes required to achieve them.

The SOMA for Autonomous Robotic Agents represents the physical and social context of everyday activities to facilitate accomplishing tasks that are trivial for humans (Beßler et al., 2021). It is based on DUL, extending their concept to different event types such as action, process, and state, the objects that participated in the activities, and the execution concept. It is worth mentioning that SOMA was intended to be used in the runtime along with the concept of narratively enabled episodic memories (NEEMs), which are comprehensive logs of raw sensor data, actuator control histories, and perception events, all semantically annotated with information about what the robot is doing and why using the terminology provided by SOMA.

The relationship between upper-level ontologies and domain ontologies is a relationship of progressive domain focalization (Sanz et al., 1999). The frameworks described in the following sections are mostly specializations of these general robotic ontologies and other foundations.

KnowRob ¹³ is a framework that provides a KB and a processing system to perform complex manipulation tasks (Tenorth and Beetz, 2009; Tenorth and Beetz, 2013). KnowRob2 ¹⁴ (Beetz et al., 2018) represents the second generation of the framework and serves as a bridge between vague instructions and specific motion commands required for task execution.

KnowRob2’s primary objective is to determine the appropriate action parametrization based on the required motion and identify the physical effects that must be achieved or avoided. For example, if the robot is asked to pick up a cup and pour out its contents, the KB retrieves the necessary action of pouring, which includes a sub-task to grasp the source container. Subsequently, the framework queries for pre-grasp and grasp poses, along with grasp force, to establish the required motion parameters.

KnowRob ontology includes a spectrum of concepts related to robots, including information about their body parts, connections, sensing and action capabilities, tasks, actions, and behavior. Objects are represented with their parts, functionalities, and configuration, while context and environment are also taken into account. Additionally, the ontology incorporates temporal predicates based on event logic and time-dependent relations.

While KnowRob’s initial ontology was based on Cyc (Lenat, 1995), which was designed to understand how the world works by representing implicit knowledge and performing human-like reasoning, Cyc has remained proprietary. OpenCyc, a public initiative related to Cyc, is no longer available. The KnowRob framework later transitioned to DUL, which was chosen for its compatibility with concepts for autonomous robots. KnowRob uses Prolog as a query and assert interface, but all perception, navigation, and manipulation actions are encoded in plans rather than Prolog queries or rules.

One of the main expansions of KnowRob is RoboEarth, a worldwide open-source platform that allows any robot with a network connection to generate, share, and reuse data (Waibel et al., 2011). It uses the principle of linked data connections through web and cloud services to speed up robot learning and adaptation in complex tasks.

RoboEarth uses an ontology to encode concepts and relations in maps and objects and a SLAM map that provides the geometry of the scene and the locations of objects with respect to the robot (Riazuelo et al., 2015). Each robot has a self-model that describes its kinematic structure and a semantic model to provide meaning to robot parts, such as a set of joints that form a gripper. This model also includes actions, their parameters, and software components, such as object recognition systems (Tenorth et al., 2012).

Another example of a KnowRob-based application is Crespo et al. (2018). In this case, the focus is on using semantic concepts to annotate a SLAM map with additional conceptualizations. This application diverges somewhat from KnowRob’s initial emphasis on robot manipulators. The work models the environment utilizing semantic concepts but specifically captures the relationships between rooms, objects, object interactions, and the utility of objects.

The detailed analysis of comparison criteria follows:

• Perception and categorization: KnowRob and RoboEarth incorporate inference mechanisms to abstract sensing information, particularly in the context of object recognition (Beetz et al., 2018; Crespo et al., 2018). As highlighted in Tenorth and Beetz (2013), inference processes information perceived from the external environment and abstracts it to the most appropriate level while retaining the connection to the original percepts. These ontologies serve as shared conceptualizations that accommodate various data types and support various forms of reasoning: effectively handling uncertainties arising from sensor noise, limited observability, hallucinated object detection, incomplete knowledge, and unreliable or inaccurate actions (Tenorth and Beetz, 2009).

This approach equips KnowRob with the ability to reason about specific physical effects that can be achieved or avoided through its motion capabilities. Although KnowRob’s KB might exhibit redundancy or inconsistency, the reasoning engine computes multiple hypotheses, subjecting them to plausibility and consistency checks and ultimately selecting the most promising parametrization. The planning component of KnowRob2 is tailored for motion planning and solving inverse kinematics problems. Tasks are dynamically assembled on the basis of the robot’s situation.

• Prediction and monitoring: KnowRob uses its ontology to represent the evolution of the state, facilitating the retrieval of semantic information and reasoning (Beetz et al., 2018). Through these heterogeneous processes, the framework can predict the most appropriate parameters for a given situation. However, the monitoring capabilities within this framework are limited to objects in the environment. Unsuccessful experiences are labeled and stored in the robot’s memory, contributing to the selection of action parameters in subsequent scenarios. The framework introduces NEEMs, allowing queries about the robot’s actions, their timing, execution details, success outcomes, the robot’s observations, and beliefs during each action. This knowledge is used primarily during the learning process.

RoboEarth and earlier versions of KnowRob rely on action recipes for execution. Before executing an action recipe, the system verifies the availability of the skills necessary for the task and orders each action to satisfy the constraints. In cases where the robot encounters difficulties in executing a recipe, it downloads additional information to enhance its capabilities (Tenorth et al., 2012). Once the plan is established, the system links robot perceptions with the abstract task description given by the action recipe. RoboEarth ensures the execution of reliable actions by actively monitoring the link between robot perceptions and actions (Waibel et al., 2011).

• Communication and coordination: RoboEarth (Tenorth et al., 2012) uses a communication module to facilitate the exchange of information with the web. This involves making web requests to upload and download data, allowing the construction and updating of the KB.

The Perception and Manipulation Knowledge (PMK) ¹⁵ framework is designed for autonomous robots, with a specific focus on complex manipulation tasks that provide semantic information about objects, types, geometrical constraints, and functionalities. In concrete terms, this framework uses knowledge to support task and motion planning (TAMP) capabilities in the manipulation domain (Diab et al., 2019).

PMK ontology is grounded in IEEE Standard 1872.2 (IEEE SA, 2022) for knowledge representation in the robotic domain, extending it to the manipulation domain by incorporating sensor-related knowledge. This extension facilitates the link between low-level perception data and high-level knowledge for comprehensive situation analysis in planning tasks. The ontological structure of PMK comprises a meta-ontology for representing generic information, an ontology schema for domain-specific knowledge, and an ontology instance to store information about objects. These layers are organized into seven classes: feature, workspace, object, actor, sensor, context reasoning, and actions. This structure is inspired by OUR-K (Lim et al., 2011), which is further described in Section 5.3).

The detailed analysis of comparison criteria follows:

• Perception and categorization: The PMK ontology incorporates RFID sensors and 2D cameras to facilitate object localization, explicitly grouping sensors to define equivalent sensing strategies. This design enables the system to dynamically select the most appropriate sensor based on the current situation. PMK establishes relationships between classes such as feature, sensor, and action. For example, it stores poses, colors, and IDs of objects obtained from images.

Failure Interpretation and Recovery in Planning and Execution (FailRecOnt) ¹⁶ is an ontology-based framework featuring contingency-based task and motion planning. This innovative system empowers robots to handle uncertainty, recover from failures, and engage in effective human–robot interactions. Grounded in the DUL ontology, it addresses failures and recovery strategies, but it also takes some concepts from CORA and SUMO for robotics and ontological foundations, respectively.

The framework identifies failures through the non-realized situation concept and proposes corresponding recovery strategies for actions. To improve the understanding of failure, the ontology models terms such as causal mechanism, location, time, and functional considerations, which facilitates a reasoning-based repair plan (Diab et al., 2020; Diab et al., 2021).

Failure ontology requires a system knowledge model in terms of tasks, roles, and object concepts. Lastly, FailRecOnt has some similarities with KnowRob; both target manipulation tasks are at least partially based on DUL and share some of the authors. Moreover, they propose using PMK as a model of the system (Diab et al., 2021).

The detailed analysis of comparison criteria follows:

• Perception and categorization: Perception in FailRecOnt is limited to action detection to detect abnormal events. For geometric information and environment categorization, the framework leverages PMK to abstract perceptual information related to the environment.

Probabilistic Logic Module (PLM) offers a framework that integrates semantic and geometric reasoning for robotic grasping (Antanas et al., 2019). Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the framework.

The primary focus of this work is on an ontology that generalizes similar object parts to semantically reason about the most probable part of an object to grasp, considering object properties and task constraints. This information is used to reduce the search space for possible final gripper poses. This acquired knowledge can also be transferred to objects within the same category.

The object ontology comprises specific objects such as cup or hammer, along with supercategories based on functionality, such as kitchen container or tool. The task ontology encodes grasping tasks with objects, such as pick and place right or pour in. Additionally, a third ontology conceptualizes object-task affordances, considering the manipulation capabilities of a two-finger gripper and the associated probability of success.

In using the ontology, high-level knowledge is combined with low-level learning based on visual shape features to enhance object categorization. Subsequently, high-level knowledge utilizes probabilistic logic to translate low-level visual perception into a more promising grasp planning strategy.

The detailed analysis of comparison criteria follows:

• Perception and categorization: This approach is based on visual perception, employing vision to identify the most suitable part of an object for grasping. The framework utilizes a low-level perception module to label visual data with semantic object parts, such as detecting the top, middle, and bottom areas of a cup and its handle. The probabilistic logic module combines this information with the affordances model.

Table 1 provides a concise comparison of the four frameworks analyzed. Although all of them address perception and categorization, decision making and planning, and reasoning, the specific aspects involved depend on the perspective. Only half of the frameworks explicitly utilize their KB for execution, prediction and monitoring, communication and coordination, and learning. In particular, integration and design are addressed exclusively by the PMK framework.

Teleological and Ontological Model for Autonomous Systems (TOMASys) ¹⁷ is a metamodel designed to consider the functional knowledge of autonomous systems, incorporating both teleological and ontological dimensions. The teleological aspect includes engineering knowledge, which represents the intentions and purposes of system designers. The ontological dimension categorizes the structure and behavior of the system.

TOMASys serves as a metamodel to ensure robust operation, focusing on mission-level resilience (Hernández et al., 2018). This metamodel relies on a functional ontology derived from the Ontology for Autonomous Systems (OASys) (Bermejo-Alonso et al., 2010), establishing connections between the robot’s architecture and its mission. The core concepts in TOMASys include functions, objectives, components, and configurations. However, it operates as a metamodel and intentionally avoids representing specific features of the operational environment, such as objects, maps, etc.

At the core of the TOMASys framework is the metacontroller. While a conventional controller closes a loop to maintain a system component’s output close to a set point, the metacontroller closes a control loop on top of a system’s functionality. This metacontroller triggers reconfiguration when the system deviates from the functional reference, allowing the robot to adapt and maintain desired behavior in the presence of failures. Explicit knowledge of mission requirements is leveraged for reconfiguration using the system’s functional specifications captured in the ontology.

In practice, TOMASys has been applied to various robots and environments, particularly for navigation tasks. Examples include its application to an underwater mine explorer robot (Aguado et al., 2021) and a mobile robot patrolling a university campus (Bozhinoski et al., 2022).

The detailed analysis of comparison criteria follows:

• Decision making and planning: In TOMASys, the metacontrol system manages decision making by adjusting parameters and configurations to address contingencies and mission deviations. It assumes the presence of a nominal controller responsible for standard decisions. In the case of failure detection or unmet mission requirements, the metacontroller selects an appropriate configuration. The planning process is integrated into the metacontrol subsystem, where reconfiguration decisions can impact the overall system plan, potentially altering parameters, components, functionalities, or even relaxing mission objectives to ensure task accomplishment.

The Ontology-based Multi-layered Robot Knowledge Framework (OMRKF) aims to integrate high-level knowledge with perception data to enhance the intelligence of a robot in its environment (Suh et al., 2007). Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the framework.

The framework is organized into knowledge boards, each representing four knowledge classes: perception, activity, model, and context. These classes are divided into three knowledge levels (high, middle, and low). Perception knowledge involves visual concepts, visual features, and numerical descriptions. Similarly, activity knowledge is classified into service, task, and behavior, while model knowledge includes space, objects, and their features. The context class is organized into high-level context, temporal context, and spatial context.

At each knowledge level, OMRKF employs three ontology layers: (a) a meta-ontology for generic knowledge, (b) an ontology schema for domain-specific knowledge, and (c) an ontology instance to ground concepts with application-specific data. The framework uses rules to define relationships between ontology layers, knowledge levels, and knowledge classes.

OMRKF facilitates the execution of sequenced behaviors by allowing the specification of high-level services and guiding the robot in recognizing objects even with incomplete knowledge. This capability enables robust object recognition, successful navigation, and inference of localization-related knowledge. Additionally, the framework provides a querying-asking interface through Prolog, enhancing the robot’s interaction capabilities.

The detailed analysis of comparison criteria follows:

• Perception and categorization: OMRKF includes perception as one of its knowledge classes, specifically addressing the numerical descriptor class in the lower-level layer. Examples of these numerical descriptors are generated by robot sensors and image processing algorithms such as Gabor filter or scale-invariant feature transform (SIFT) (Suh et al., 2007).

The Smart and Networked Underwater Robots in Cooperation Meshes (SWARMs) ontology addresses information heterogeneity and facilitates a shared understanding among robots in the context of maritime or underwater missions (Li et al., 2017). Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the ontology.

SWARMs leverages the probabilistic ontology PR-OWL ¹⁸ to annotate the uncertainty of the context based on the multi-entity Bayesian network (MEBN) theory (Laskey, 2008). This allows SWARMs to perform hybrid reasoning on (i) the information exchanged between robots and (ii) environmental uncertainty.

SWARMs establishes a core ontology to interrelate several domain-specific ontologies. The core ontology manages entities, objects, and infrastructures. These ontologies include:

– Mission and Planning Ontology: Provides a general representation of the entire mission and the associated planning procedures.

Li et al. (2017) present an example using SWARMa to monitor chemical pollution based on a probability distribution. SWARMs incorporates this model into the ontology and uses MEBN to deduce the emergency level of the polluted sea region.

The detailed analysis of comparison criteria follows:

• Perception and categorization: The SWARMs ontology provides a shared framework to represent the underwater environment. For this reason, it contains classes on sensors and the main concepts of the environment to understand it through its properties, such as water salinity, conductivity, temperature, and currents. Uncertainty reasoning is critical for the categorization of sensor information, especially in harsh maritime and underwater environments.

The Robot Task Planning Ontology (RTPO) Sun et al. (2019b) is an effective knowledge representation framework for robot task planning. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the ontology.

Designed to accommodate temporal, spatial, continuous, and discrete information, RTPO prioritizes scalability and responsiveness to ensure practicality in task planning. The ontology comprises three main components: robot, environment, and task.

– Robot Ontology: Comprises hardware and software details, location information, dynamic data, and more. Sensors are explicitly modeled as a subclass of hardware, specifying the measurable aspects of the environment. In an experimental context, the robot’s perception is limited to obstacles.

The detailed analysis of comparison criteria follows:

• Perception and categorization: The robot ontology in RTPO considers sensors as a subclass of hardware, specifying measurable aspects of the environment. In the experiment discussed, the robot’s perception is limited to detecting obstacles.

The task planning process is realized by matching the execution preconditions of atomic actions and their effects on the environment from the initial state to the goal state. With this approach, the task planning module can also be executed by changing the input tasks while considering the present action resources. The plans generated by the task planning algorithm are added to the ontology to improve the efficiency of task planning if the same task needs to be planned again.

• Reasoning: The reasoning process in RTPO involves updating and storing the knowledge within the ontology. The scalability of robot knowledge aims to enhance the efficiency of reasoning. Experimental scenarios involving the addition of elements to the indoor environment and corresponding KB instances demonstrate changes in consumed time, particularly affecting knowledge query speed. The authors show real-time performance in an application that involves 52,000 individuals, although the impact on the planning process is not explicitly detailed.

Burroughes and Gao (2016) present an architectural solution to address limitations in autonomous software and GNC structures designed for extraterrestrial planetary exploration rovers. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the framework.

This framework uses an ontology to facilitate the autonomous reconfiguration of mission goals, software architecture, software components, and the control of hardware components during runtime. To manage complexity, the self-reconfiguration ontology is organized into modules. The base ontology serves as an upper ontology, including modules that delineate logic, numeric aspects, temporality, fuzziness, confidences, processes, and block diagrams. Additional modules describe the functions of software, hardware, and the environment.

The primary focus of this framework is reconfiguration. Once the ontology manager detects an undesired state change through the monitoring process, it activates a safety mode. In this safety mode, the system can execute either a reactive plan or create a new plan based on inferences drawn from the new situation.

Following the replanning process, new mission goals are established, allowing the robot to exit safety mode and resume normal operation. The framework specifically aims to minimize odometry errors and ensure safety during travel. To achieve this objective, the connective architecture and processes, such as navigation, localization, control, mapping, as well as sensors and the locomotion system, can be fine-tuned and optimized according to the environment and faulty hardware conditions. For example, certain methods may prioritize tolerance to sensor noise over absolute accuracy in localization.

The detailed analysis of comparison criteria follows:

• Decision making and planning: Burroughes and Gao (2016) integrate a reactive approach with a deliberative layer for quick responses. Precalculated responses are prepared for likely changes, but in the absence of options, the rational agent resorts to deliberative techniques. Ontologies and the PDDL are used for knowledge representation and planning, respectively. Actions correspond to specific configurations of services, with plans defined using the initial state of the world, the goal state, the resources of the system, the safety criteria, and the rules. The approach employs a pruning algorithm to reduce possible actions and planning space.

Chandra and Rocha (2016) present a framework for collaborative context awareness using an ontology that comprises high-level aspects of urban search and rescue (USAR) missions. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the framework.

The framework serves as an efficient knowledge-sharing platform to represent and correlate various mission concepts, including those related to agents and their capabilities, scenarios, and teams. The ontology facilitates the sharing of homogeneous knowledge among all robots and supports human–robot collaboration by reasoning based on rules provided by humans.

The detailed analysis of comparison criteria follows:

• Perception and categorization: The system employs perceived and pre-processed information gathered from its own sensors, as well as from other agents, such as humans and robots, to continually update its KB. Specifically, it focuses on entities within context classes and their relationships, including information about smoke levels, visibility, location, and temperature.

Huang et al. (2019) use ontologies for situation assessment and decision making in the context of autonomous vehicles navigating urban environments. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the ontologies.

The KB integrates vehicle information, storing details of the permissible directions a vehicle can take. Furthermore, the KB includes representations of both static and dynamic obstacles, as well as relevant information on the characteristics of roads, distinguishing between highways and urban roads. The ontology incorporates specific scenarios that the autonomous vehicle could encounter on the road, such as proximity to an intersection, traversing a bridge, or executing a U-turn. The autonomous driving system leverages this KB to assess the current driving scenario, facilitating well-informed decisions on whether to maintain its current trajectory or initiate a lane change.

The detailed analysis of comparison criteria follows:

• Decision making and planning: The decision-making process is tied to situation assessment, employing a methodology that evaluates the safety of the surroundings of the vehicle. This involves characterizing the regions around the car and assigning a binary safety value based on the presence of obstacles.

If a region is considered safe, the vehicle continues in its current lane; otherwise, a lane change is initiated. This decision-making step incorporates statistical indicators that account for velocity. Moreover, the model considers legitimacy by adhering to traffic rules during lane changes and reasonableness by querying the global planning path to identify the next road segment or lane, particularly when approaching intersections. This strategic approach prevents the system from optimizing locally at the expense of compromising the overarching global plan.

• Reasoning: The reasoner is used to generate behavioral decisions. It employs rules derived from traffic regulations and driving experiences, taking into account various factors that influence the road, such as speed limits, traffic lights, and the presence of surrounding obstacles.

Sun et al. (2019a) introduce an ontology tailored for search and rescue (SAR), enhancing the decision-making capabilities of robots in complex and unpredictable scenarios. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the ontology. The SAR ontology comprises three interlinked subontologies: an entity ontology, an environment ontology, and a task ontology.

– The entity ontology includes various types of robots, including ground, underwater, and air robots. It delineates their constituent parts and specifications in terms of hardware and software aspects.

The detailed analysis of comparison criteria follows:

• Perception and categorization: The SAR framework focuses on SLAM and autonomous navigation. The authors use a semantic map, which facilitates the recognition and localization of objects. The acquired information dynamically updates the environment ontology, establishing connections between the SLAM map and the semantic details. Bayesian reasoning enhances the precision of victim positioning, while QR code scanning streamlines the acquisition of semantic information, such as vital signs (e.g., heart rate and blood pressure), in disaster rescue scenarios.

Tables 2 and 3 summarize studies in the navigation domain. All frameworks address decision making, planning, and reasoning, mainly to organize tasks. Most of the works focus on perceiving and categorizing measurable aspects of the environment, along with communication and coordination among different elements and systems. TOMASys, the planetary rover scenario, and the SAR scenario explicitly handle prediction, monitoring, and execution. The use of engineering knowledge—interaction and design—is explicitly addressed in TOMASys, the planetary rover scenario, and the autonomous vehicle scenario. Learning is applied only in RTPO to improve the plan generation process.

The Ontology-based Unified Robot Knowledge (OUR-K) framework for service robots (Lim et al., 2011) consists of five knowledge classes: features, objects, spaces, contexts, and actions. It takes OMRKF, the concept of layer division, from its predecessor. Although OMRKF was originally evaluated in navigation domains, OUR-K extends its application to social robots operating within domestic environments. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the framework.

A notable feature of OUR-K is its ability to perform tasks even when provided with incomplete information. The framework incorporates a knowledge description of both the robot and its environment, employing algorithms for knowledge association. This involves logic, Bayesian inference, and heuristics; however, logical inference in OUR-K is specifically limited to performing associations between classes and ontological levels, with no indication of alternative inference mechanisms in the literature.

OUR-K includes mechanisms for object recognition, context modeling, task planning, space representation, and navigation. Regarding action representation, its descriptions are simpler and do not contemplate processes as did its predecessor, OMRKF.

Diab et al. (2018) extend OUR-K with a physics-based manipulation ontology to address the challenges that a motion planner might encounter. An actor class is introduced within the knowledge classes to describe the working constraints of the robot, enhancing the planner’s capability to handle interaction dynamics. In addition, the authors propose a prediction mechanism for the entire OUR-K framework. Instead of relying on inferences, a semantic map is generated for categorizing and assigning manipulation constraints, using reasoning based on logical axioms. This approach is evaluated in the context of a specific manipulation task, where a robot serves a liquid drink contained in a can.

The detailed analysis of comparison criteria follows:

• Perception and categorization: OUR-K follows the same approach as OMRKF. Perception is abstracted and stored in the feature, space, and object classes. The feature class defines the same knowledge level as the OMRKF. The space class defines a topological map in the middle level and a semantic map in the higher one. The object class middle level includes the object name and function, whereas the top layer defines generic information and relationships among objects; for example, a cup is a type of container.

The OpenRobots Common Sense Ontology (ORO) (Lemaignan et al., 2010) establishes a knowledge processing framework and a common sense ontology tailored for facilitating semantic-rich human–robot interaction environments ¹⁹ . This approach focuses on conceptualization but offers flexibility for implementing other cognitive functions simultaneously, such as object recognition, task planning, or reasoning. Cooperation is a crucial aspect in ORO, as it targets human–robot interactions.

ORO ontology is based on OpenCyc. The authors specify in the project Wiki ²⁰ that it shares most of its concepts with the first version of the KnowRob ontology. It includes categories, such as spatial thing or action, and more concrete concepts, such as event or book. The authors evaluated their approach by showing robot objects and asking humans about their properties.

The detailed analysis of comparison criteria follows:

• Perception and categorization: The authors use several algorithms, such as common ancestors or calculating the best discriminant to categorize perceptions. Discrimination is an important element in human–robot collaboration. For example, if a user asks the robot to bring the bottle and two bottles are available, it can use its ability to discriminate to select the best option.

Each agent has an alternative cognitive model for the other agents with whom it has interacted. When ORO identifies a new agent, it automatically creates a new separate model that can be shared with others. This feature enables the storage and reasoning of different and potentially globally inconsistent models of the world.

Ontology for Collaborative Robotics and Adaptation (OCRA) ²¹ , as described by Olivares-Alarcos et al. (2022), is a specialized ontology designed to represent relevant knowledge in collaborative scenarios, facilitating plan adaptation. Based on KnowRob (Tenorth and Beetz, 2009; Beetz et al., 2018) and its upper ontology, DUL, OCRA shares the support of its predecessor for the temporal history of the KB through episodic memories.

This work primarily employs FOL definitions to establish a foundation for reliable, collaborative robots, with the aim of enhancing the interoperability and reusability of terminology within this domain. The ontology serves as a tool for the robot to address competency questions, such as identifying ongoing collaborations, understanding current plans and goals, determining the agents involved, and assessing plans before and after adaptation.

This ontology has been qualitatively validated for human–robot cooperation, sharing the task of filling the compartments of a tray. The main asset of OCRA, from our perspective, is its explicit representation of collaboration requirements, including safety considerations and a measure of risks, with the objective of effective plan adaptation.

The detailed analysis of comparison criteria follows:

• Decision making and planning: OCRA uses ontological knowledge for dynamic plan adaptation during runtime. For example, if the robot had a plan to fill a certain compartment and it is not empty, it adapts to fill another empty compartment. The ontology also stores the adaptation triggers.

Intelligent Service Robot Ontology (ISRO) (Chang et al., 2020) introduces an ontology-based model tailored for human–robot interaction. The practical application of this approach is demonstrated through the implementation of a social robot that functions as a medical receptionist in a hospital setting. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the ontology.

ISRO serves as an abstract knowledge management system designed to comprehend information about agents, encompassing both users and robots, as well as their environment. The ontology establishes connections between user knowledge and the robot’s actions and behaviors, incorporating considerations for the spatial and temporal environment. This integration is facilitated through the Artificial Robot Brain Intelligence (ARBI) framework, which includes a task planner, a context reasoner, and a knowledge manager.

ISRO is not limited to a specific domain because it provides a high-level scheme for dynamic generation and management of information. The ontology is divided into four sub-models:

– User ontology to store profile information.

The detailed analysis of comparison criteria follows:

• Perception and categorization: The framework uses sensor information to recognize the user involved, their face, gender, and age, and recognize their state, that is, the user’s expression, in its application. It also senses the robot’s pose to guide the user to the medical department. Additionally, the framework is also prepared to handle information related to spaces and objects that was not defined in the experiment.

Ji et al. (2012) provide a flexible framework for service robots using the automatic planner Stanford Research Institute Problem Solver (STRIPS). Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the framework.

The ontology represents two main types of information, environmental description and primitive robot actions, which handle spatial uncertainties of particular objects and the primitive actions that the robot can perform. Actions include four main attributes to comply with STRIPS: precondition, postcondition, input, and result. The planning process is optimized using a recursive back-trace search method and knowledge information to limit the search space. The framework is applied to the Care-O-Bot agent in a scenario in which it shall get a milk box.

The detailed analysis of comparison criteria follows:

• Perception and categorization: Symbol grounding is crucial for Ji et al. (2012), as it bridges the gap between abstract planning and actual robot sensing and actuation. For example, a task of moving a table involves moving the robot near the table. The symbol grounding specifies exactly where “near the table” is.

Table 4 depicts the main aspects of each capability in the social applications reviewed. All capabilities incorporate reasoning using knowledge from the ontology, with the majority also utilizing it for perception and categorization, decision making and planning, and execution. Communication and coordination are only present in two of the five analyzed works. Prediction and monitoring, interaction and design, and learning are the least addressed capabilities across these applications.

Robot control for Skilled Execution of Tasks (ROSETTA) constitutes an ontology for robots performing manufacturing tasks ²² . Its origins can be traced to the European projects SIARAS, RoSta, and ROSETTA (Stenmark and Malec, 2015). The core of the ontology is mostly focused on robot devices and their skills. It relies on a component called the knowledge integration framework (KIF) to provide services for robotic ontologies and data repositories (Stenmark and Malec, 2015). Note that this should not be confused with the knowledge interchange format, which is a syntax for FOL. KIF acts as an interface to users. They can specify the task by partially ordering the subgoals in an assembly tree; the framework then establishes the action planning and its schedule with limited resources. KIF also provides an execution structure that generates state machines from skill descriptions and their constraints.

Hoebert et al. (2021) use ROSETTA ontology to control the assembly process of a variety of electronic devices. This work focuses on a world model that represents robotic devices and the skills of ROSETTA. It also relies on the boundary representation (BREP) ontology (Perzylo et al., 2015) to semantically encode geometric entities. This work uses the ontology to conceptualize objects and their properties in the environment. It defines a product as a hierarchy of sub-assemblies and parts. Requirements are also included to determine the correct automated manufacturing.

Merdan et al. (2019) also employ the ROSETTA ontology to control an industrial assembly process, in particular, a pallet transport system and the control of an industrial robot. In this case, it uses the ROSETTA and BREP ontologies to conceptualize the robotic system (e.g., skills, properties, constraints, etc.), the product model (e.g., parts, geometries, assembly orientation, etc.), and the manufacturing infrastructure (e.g., product, storage, sensors, etc.). Rosetta has been defined in OWL.

The detailed analysis of comparison criteria follows:

Perception and categorization: The ROSETTA application by Hoebert et al. (2021) provides an object recognition module that links perception data with geometric features represented in the KB.

Decision making and planning: The KIF executor of the ROSETTA ontology serves as the planning mechanism. It transforms an assembly graph into a sequence of operations with preconditions and postconditions, subsequently translated into a task state machine. Both Merdan et al. (2019) and Hoebert et al. (2021) approach decision making as a plan generator, utilizing PDDL. Hoebert et al. (2021) provide a generator that extracts information from the ontology to produce the required PDDL files for planning. Merdan et al. (2019) also translate the semantic model, that is, states and actions, into domain and problem files through templates.

Reasoning: ROSETTA reasoning is enabled by the KIF server (Stenmark and Malec, 2015). It allows the user to download and upload libraries with object descriptions, task specifications, and skills. The KIF reasoner assists robot programming by retrieving information about tools, sensors, objects, object properties, etc. Hoebert et al. (2021) take advantage of the ROSETTA ontology to select the individual actions and the equipment required to manufacture a product part.

Execution: KIF uses state machines to execute the skills defined in the ROSETTA ontology (Stenmark and Malec, 2015). However, Hoebert et al. (2021) and Merdan et al. (2019) use the execution of atomic actions through PDDL commands. With this approach, as discussed by Hoebert et al. (2021), the missing standardization of the meta-level concepts can cause difficulties when integrating existing ontologies and reduce the scalability of the approach. A comparison of the performance using several PDDL planners is discussed in this article. In terms of scalability, Merdan et al. (2019) use a central database server to support a high-performance integration of multiple homogeneous data sources to integrate the ontology with other knowledge sources.

Communication and coordination: The ROSETTA-based framework by Merdan et al. (2019) aims to communicate between robots and entities in the external production environment. Additionally, being a component-based approach, it provides infrastructure for intercomponent communication.

Interaction and design: ROSETTA provides an engineering specification of workspace objects, skills, and tasks from libraries (Stenmark and Malec, 2015). It acts as a database for all the information present in the object and in the environment. Hoebert et al. (2021) use these design models to establish and verify the manufacturing process, tailoring the requirements to parameters such as assembly operation type, manufacturing constraints, material used, and piece dimensions.

Bernardo et al. (2018) define an ontology-based approach for an industrial robotic application focused on inserting up to 56 small pins (sealants) into a harness box terminal specifically tailored for the automotive industry. Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the approach.

The KB is built upon CORA (Prestes et al., 2013) and provides specifications for the robot, the machine vision system, and the tasks required for seamless operation. Task sequences are used to execute the plan, but the machine vision system also uses them to inspect whether the sealants were correctly inserted into the connecting boxes. If not, the system prioritizes applying the sealant in the faulty position(s) in the connecting box.

The detailed analysis of comparison criteria follows:

• Perception and categorization: This approach uses vision to inspect robot operations, specifically to verify the correct insertion of sealants in connecting boxes. Image processing is applied for this inspection, although the information is not categorized in the ontology. The framework also utilizes proximity, position, and fiber optic sensors to facilitate robot operation.

Borgo et al. (2019) extend the DOLCE ontology from a broad perspective. They validated their approach using a real pilot plant, a reconfigurable manufacturing system designed for recycling printed circuit boards (PCBs). Specific details about the KB, such as the source files, are not publicly available, so the information provided here is derived from articles about the ontology. The ontological framework aims to store knowledge about fundamental assumptions and identify the current scenario, including the presence and location of objects, executed actions, responsible agents, changes occurring, etc.

The KB created by this approach seeks to integrate various perspectives within the enterprise, covering intelligent agents, engineering activities, and management activities. The framework is structured in a deliberative layer to synthesize the actions needed to achieve a goal, an executive and monitoring layer to verify the actions, and a mechatronic system that determines the capabilities of the agent.

• Decision making and planning: The updated KB uses an abstraction of the device to be controlled, the environment parameters, and the execution constraints to generate a timeline-based planning model. This model provides the atomic operations that a transportation module can perform, depending on its components and the available collaborators. The timeline incorporates temporal flexibility, allowing for relaxed start and end times. The resulting plan represents the potential evolution of the relevant feature. The framework also supports replanning in case of plan execution failure, generating a new plan based on the current state of the mechatronic system.

Table 5 provides a summary of three frameworks applied to industrial scenarios. ROSETTA encompasses all capabilities except prediction, monitoring, and learning. Similarly, the adaptive agent for manufacturing domains includes all capabilities except perception, categorization, and learning. Lastly, the remaining industrial scenario only covers decision making, planning, and reasoning, which were required to be included in the review, and perception and categorization.

Table 6 depicts other aspects of the frameworks studied, such as the languages used, the use of temporal conceptualizations, and the foundational ontologies.

Almost all works use OWL or a combination of OWL and Prolog as the KB language. For planning, most of them use geometric planners, such as PLM (Antanas et al., 2019), along with task planners. In some works, task planners are custom made, such as the service robot from Ji et al. (2012). However, in general, task planners are based on ontological queries, such as the SAR scenario (Sun et al., 2019a), the industrial application (Bernardo et al., 2018), and the adaptive manufacturing of Borgo et al. (2019). The planetary rover (Burroughes and Gao, 2016) and ROSETTA (Stenmark and Malec, 2015) use PDDL directly, while RTPO (Sun et al., 2019b) and SWARMs (Li et al., 2017) use a custom approach similar to PDDL.

Some studies use temporal conceptualizations to address planning dynamics. Most of them rely on an ordered sequence of actions. However, KnowRob (Tenorth and Beetz, 2009; Beetz et al., 2018), OCRA (Olivares-Alarcos et al., 2022), and the adaptive manufacturing approach use time intervals to recall a time frame in which the action is executed. OMRKF (Suh et al., 2007) and OUR-K (Lim et al., 2011) include time as part of the context ontology, and the planetary rover (Burroughes and Gao, 2016) and ISRO (Chang et al., 2020) conceptualize time in their KBs.