Autonomy in the Age of Knowledge Graphs: Vision and Challenges

The first dimension of autonomy that we consider is the autonomy of an agent in relation to its environment. According to Castelfranchi and Falcone [17], this relates to agents being able to act in an environment based on their perception of this environment while maintaining autonomy from the perceived stimuli and from the environment itself. This is illustrated in [17] with with a billiard ball – which is not autonomous with respect to its environment – and it is argued that agents who indeed are independent from their environment instead feature teleonomic behaviour, i.e., that they actively perceive and interpret their environment and the effects of their actions, that they orient themselves towards the input, and that their behaviour depends on internal states with their own evolution principles – rather than merely receiving an environmental “force” [17] as the billiard ball does. With respect to the environment dimension, the autonomy of an agent in an environment is increased when this agent gains knowledge about artifacts that are present (or likely present) in that environment. This includes knowing how it could successfully carry out low-level interactions with these artifacts – that is, information about the user interface or API of the artifact; it furthermore encompasses knowledge about how to discover artifacts in an environment (e.g., through an artifact registry or through broadcasting), and knowledge that permits an agent to establish that a specific artifact is relevant to an agent’s course of action – i.e., that using the artifact would be beneficial for the agent to achieve its goal.

The benefits of creating agents that are more autonomous from their environment are sought after across many fields: In software engineering, this relates to the creation of software systems that emphasise the non-functional properties of adaptability and resilience – that is, systems that are still able to perform their core function even if their environment changes. These properties are typically – and, often, consciously – traded-off against performance: In a system that emphasises performance over adaptability, there is a high incentive to tightly couple even low-level interactions and to forgo run-time discovery capabilities by rather hard-coding device addresses and service APIs. Increasing the autonomy of a system with respect to its environment thus requires, explicitly or implicitly, to implement such systems in a way that does not couple their function to the interfaces of specific artifacts in the environment (e.g., by hard-coding a specific HTTP request); one layer above such interface coupling, the same goal motivates the creation of systems whose functions are not even coupled to specific artifacts – such systems hence need to be enabled to discover artifacts at run time [80], and then require an (internal or external) way to evaluate whether using these artifacts would contribute to the agents achieving their goals (e.g., a planner). The same fundamental principle is mirrored in education, a seemingly vastly different domain: In a similar way as it might be desirable to implement a software system that has enough knowledge to reason about the usage of specific artifacts while not being hard-coded to use them, we typically teach children in a way that increases their autonomy from specific environmental artifacts. To illustrate this, it is perfectly feasible to instruct a child – at the interface level – to use a specific means of public transport at a specific time and location to reach a specific destination – “To get to your Judo class, you need to always use bus line #2 at 5:37pm”. However, parents who emphasise autonomy of their child in relation to the environment will rather opt for explaining to the child – at the knowledge level – how a public transport system is used in general, which involves knowledge about public transportation, ticketing, routes, schedules, and possibly even how to read maps or use digital tools for navigation. This provides the child with information about how to navigate their environment effectively even when the environment is dynamic, and hence increases the child’s autonomy in the environment – e.g., when bus #2 is cancelled.

We argue that KGs are highly suited to support this type of autonomy of agents from their environment, i.e., with respect to the specifics of how to interact with things, by furnishing them with information about the artifacts in the environment and where this information is compatible with the information in the agents’ own knowledge base – in this case, this knowledge can be readily integrated and processed by the agent. Information about how to interact with (heterogeneous) artifacts in an environment would build on top of a basic (homogeneous) interaction ability; on the Web, this ability is hypermedia-based interaction, i.e., that an agent is able to use hypermedia controls (including unparameterised hyperlinks as well as forms). The first steps towards this have already been made, and have even been standardised by the World Wide Web Consortium (W3C): the goal of the W3C Web of Things Thing Description⁵⁵5https://www.w3.org/TR/wot-thing-description11/ (W3C WoT TD) is precisely to permit agents that understand this standard to use any W3C WoT TD-described device they encounter in their environment. This is accomplished by conveying details about the usage of the interface of this device at run time. This is not only academically interesting, but has convincing practical application: For example, it means that a W3C WoT TD-described sensor can be upgraded to a better model without the requirement to update clients of the sensor, as long as the two sensors’ outputs are compatible with respect to the clients’ common knowledge. Further, while W3C WoT TDs today remain limited, they may be generalised to what is referred to as Artifact Profiles: “structured data describing the artifact through signifiers and general (domain- and application-specific) metadata” [87]. Such profiles hence describe features of an environmental artifact beyond its interface, and can be used by agents at run time to determine whether an artifact is interesting to them in their current situation – again increasing the agents’ environmental autonomy. It has also been proposed that such profiles could be brokered to agents based on the current situation and abilities of the agent; this information is in this case shared in the form of an Agent Profile [54, 87]. In this scope, the SHACL vocabulary can be used to define behaviour specifications that describe to agents how to use an artifact (e.g., how to adjust a lamp’s brightness through an HTTP API) [83].

Regarding the second dimension of autonomy – social autonomy – the autonomy of an agent in its environment is also increased when this agent has sufficient knowledge about other agents that are present in this environment, which should not depend on the hard-wiring of agents’ social capabilities (either directly or through their environment). Similar to autonomy from the environment, we also consider social autonomy when training autonomous agents – this is less prevalent in software engineering, where software programs (themselves) are not yet being considered primarily as social entities, but it is highly visible in human education. In fact, this includes core tenets of education itself, including the ability to speak (multiple languages) and read, as well as neurological maturation. For example, the development of Theory of Mind (cf. [72]) in infants and, with it, the knowledge that one’s own beliefs, desires, emotions, and thoughts are different from those of others, is considered crucial for successful everyday social interactions among heterogeneous humans.

Following Castelfranchi and Falcone [17], we may distinguish between autonomy as independence – the ability of an agent to reach its goals without the help of other specific agents – and autonomy in delegation, i.e., an agent’s ability to achieve delegated tasks without detailed imperative instructions (colloquially: without micro-management). We argue that both of these sub-dimensions of autonomy can benefit from KGs in similar ways, as they require similar technical capabilities:

1. i)

   self-sufficiency is a key requirement that provides the basic abilities necessary to achieve a goal or complete a delegated task, irrespective of the particular environment and the agents therein; and

2. ii)

   in either case, generally applicable social communication abilities enable the utilisation of other agents’ abilities and resources in order to complete the goal or task given the current specifics of the environment.

In Section 3.1, we discussed how interface specifications can avoid the hard-wiring of two specific system interfaces to each other (including the tight coupling of agents to environment specifics), and introduced that this supports autonomy – specifically, by facilitating adaptability and resilience through dynamic (re)composability. In the same way, agents may avoid tight coupling with other agents by utilising knowledge about their social environment together with generic abilities to discover new knowledge by means of interacting with other agents. Concretely, and mirroring the environment dimension, such agents would not be hard-coded to interact with specific other agents in specific ways, but would rather be equipped with more broadly applicable knowledge about other agents – this knowledge might include that other agents might have specific abilities, that the agent might consult with them or delegate tasks to them, and that the likelihood of other agents to engage in an interaction is determined through social relationships between agents. An agent that is equipped with this knowledge – much like a human who understands social relationships – is much better suited to navigate a social situation, and hence features increased social autonomy.

KGs are very well suited to implement this type of knowledge-based programming, as they can augment an agent’s interactive capabilities and hence decrease its dependency on other specific agents. As in the case of autonomy from the environment, autonomous agents in this context may also benefit from both accessing external KGs and maintaining their own ones. Agents may then use internal KGs to maintain knowledge that is specific to the agent and knowledge that is sufficiently important for being logically grouped with a given agent, thus facilitating this agent’s independence. Some of the internal knowledge may be strictly private, while other knowledge can be shared with other agents, or be entirely public. For example, the agent may maintain lookup services of its own capabilities, used for provisioning and planning, many of which can be shared publicly, which facilitates collaboration with other agents. In contrast, knowledge about other specific agents that the agent maintains, e.g., for the purpose of strategic reasoning and negotiation, might remain private to start with. It may be disclosed eventually however, if by doing so, this is useful in the context of a specific objective, e.g., during a multi-agent dialogue. Finally, privacy-relevant knowledge such as health information about a natural person will typically remain private, but may be shared with authorised agents on a need-to-know basis. When KGs are applied within an agent, they can be seen as knowledge representation tools that are used to persist an agent’s belief base, or snapshots thereof, considering that it may not be feasible to capture all belief revision operations that are executed in the course of the agent’s reasoning cycle. In case external KGs are provided as part of the environment, agents might use them for stigmergy, i.e., for coordination through environmental artefacts that are created or manipulated by the agents [73, 76] (see Section 3.1). This can facilitate multi-agent coordination and relax the coupling between agent instances, thus facilitating resilience and ability to respond to change.

Going beyond the interactions of agents where agent goals are generally aligned, organisations provide a deontic framework for agents to operate in by specifying joint objectives, norms, and policies. On the meta-level, they also permit agents to participate in organisational governance. The idea of virtual organisations as a means to facilitate autonomous (multi-agent) systems is a research direction that emerged at the turn of the century [18, 32, 67] and has since continued as a vibrant area of research [44]. While applications of multi-agent organisations in the sense of the holistic academic vision are nascent, most real-world organisations have by now some (often crucially important) form of electronic representation, and substantial efforts are incurred to ensure organisational compliance, e.g., with respect to privacy regulations [23].

In the context of organisational governance, KGs already exist in industry organisations for providing shared domain models that can be used across software systems and application scenarios, thus facilitating interoperability and centralising some of the knowledge maintenance efforts [40]. Beyond such current scenarios, at the organisational level, KGs are well-suited to enable agents to adjust to and switch between organisational contexts with greater independence, covering a broad range of concerns in the context of organisational autonomy. Goal-driven agents may act in a system containing several organisations, and interact with these organisations considering agent-internal goals. We claim that an agent that is aware of organisations (in general) and can hence better understand specific organisations it encounters at run time is able to reach higher autonomy. Such an agent would, for instance, know that another agent is deontically bound to performing a specific service, and that it may hence interact with that agent towards fulfilling a goal that requires this service. This is again similar to human interactions: Humans learn that they may trust a bus driver to share accurate information about bus routes. Crucially, they further know that this trust is modulated not by the bus driver (as an autonomous agent) but it is due to the bus driver being bound by the organisational principles of the public transport company they work for. In addition to interacting with organisations, agents might join organisations to then achieve an alignment of potentially conflicting knowledge or goals. That is, an agent’s decision to join a specific organisation and to adopt a particular organisational role can be informed by knowledge that the agent obtains about the organisation, prior to joining.

Technically, this may be enabled through agents exchanging knowledge about organisations through KGs directly, or by making this information discoverable through environmental artifacts; the latter case also can be extended to permit (sufficiently authorised) agents to modify organisational knowledge at run time and hence allows groups of agents to evolve organisations together. Furthermore, KGs can provide agents that join an organisation with fundamental knowledge that is required for successful (inter-)action in the context of this organisation; i.e., knowledge about the available roles and objectives, and even knowledge that helps an agent decide about whether it should commit to a specific role in the organisation. To this end, KGs can provide deontic specifications of the organisation’s objectives, as well as of the policies and norms that agents must and should, respectively, comply with in order to reach the organisational objectives or agent-specific goals. Since these goals may conflict with organisational objectives and with the goals of other agents, knowledge about sanctioning mechanisms may also be shared, which again leads to an increase in an agent’s autonomy with respect to the organisation. Equipped with this knowledge, it can now make an informed decision about whether or not to break norms that are set by the organisation. Here, one can view the interaction of agents with the organisation (or with each other through the organisation) from a governance perspective: the organisation governs agent behaviour through norms and policies, and the organisation’s agents govern the organisation, e.g., by changing organisational policies or by gradually evolving norms. Then, KGs can provide a transparent representation not only of the objectives, policies, and norms, but also about the meta-framework that governs them, thus enabling agents to affect organisational change in a structured manner. With respect to technologies, constraints on KGs – which can be represented using languages such as SHACL – can play an important role for organisations of autonomous agents. For example: constraints that are to be satisfied in the future may model organisational goals; constraints that have to be satisfied amount to policies; and constraints that should typically be satisfied may represent norms.

So far, we have discussed how knowledge can support autonomy from an external perspective: an agent is typically situated in an environment (see Section 3.1) and interacts with artifacts and other agents (see Section 3.2), and potentially as part of an organisation (see Section 3.3). Across these dimensions, we argued that a formal, explicit, knowledge-level representation of relevant entities allows the agent to reason and decide about its interactions with the rest of the system.

We now finally survey the agent’s internal perspective, where autonomy relates to the agent’s freedom of selecting the goals it is working towards, and of executing a course of action for achieving selected goals. Following Castelfranchi and Falcone [17], an agent has goal autonomy if: i) it is endowed with its own goals (e.g., at design time); or ii) it adopts goals received from other agents only when the adopted goals enable some of its own goals. The latter relates to the discussion in Section 3.2 about autonomy in delegation, but from an internal perspective that focuses on the receiving agent’s ability to select its own goals. Goal autonomy hence is compatible with social ability: If an agent interacts with other agents and can adopt goals from them, then the agent is susceptible to influence – but the agent still preserves its goal autonomy if it can reason about the relations between candidate goals and its own goals, and if it can select specific goals to be adopted based on this reasoning. We argue that, similar to how KGs enable an agent to achieve autonomy with respect to the environment, other agents, and organisations, KGs may also help an agent in this very reasoning process, and hence help it achieve goal autonomy. KGs would do this by representing explicitly or implicitly relations between candidate goals and the agent’s own goals; they may also represent commonsense knowledge (e.g., see [42]), thereby providing a foundation for the agent’s reasoning about goal adoption.

After selecting a goal, an agent will typically strive to execute a course of action towards achieving that goal. To do this, it requires procedural knowledge, which can take various forms: For example, procedural knowledge can be defined in terms of plans that specify sequences of actions and are programmed by a developer. Conversely, within a marketplace, an auctioneer may need to declare formally the dimensions along which the auction protocols may vary (ascending or descending, sealed-bid or open-cry, single-shot or continuous, first or second price, etc.) to allow agents to reason whether or not they should engage in the auction, as well as providing evidence to bidders of the auction outcome [6]. In the case of a reinforcement learning agent, procedural knowledge takes the form of a policy that is learned through repeated interactions with the environment. In other systems, from a practical/industry-oriented perspective, procedural knowledge could be represented through workflows, processes, or standard procedures [5, 74]. Agents’ goal achievement autonomy – that is, their executive autonomy – is increased if agents are able to acquire procedural knowledge at run time. However, the means to represent procedural knowledge depend on the internals of the agents using the knowledge. In some cases, KGs might provide a suitable solution for representing such knowledge themselves [88]. In the more general case, KGs may provide a uniform way of describing and linking to procedural knowledge such that agents can then acquire the knowledge in the context of their operation. We hence propose that KGs may not only directly enable executive autonomy, but that they may – across many scenarios – provide a uniform interface to procedural knowledge that itself is represented in diverse formats.

We propose that KGs provide a solid foundation for supporting autonomy at different levels, including environmental, social, and deontic aspects as well as an agent’s internal goal-selection and execution. However, this knowledge is neither static nor immutable [24]: In the case of knowledge regarding the agents’ environment, this may change over the course of time, and represent variations in the conditions and stimuli accessible to the agents. Individual and communal knowledge evolves according to the changes in the internal state of the agents, as well as according to the different interactions that they may engage in. Moreover, this knowledge is not simply available for querying as a set of read-only resources, but it is actually subject to modifications by the very autonomous agents that use it. Using KGs to represent this knowledge caters to these requirements: KGs – and the contained entities – feature unique identifiers and can hence be easily and unambiguously addressed and, if need be, versioned, in centralised as well as decentralised settings.

The way these KGs are constructed may differ in terms of approach and/or underlying technologies [89]: Individual autonomous agents might develop their own knowledge independently of their peers, where they will in the case of KGs typically use at least a common standard ontology or controlled vocabulary, in order to allow interoperability at the time of inter-agent interactions. On the other hand, if agents are limited to updating existing knowledge with new information (e.g., in the case of new observations from a sensor), the knowledge structure does not change often, but the observations may rapidly become stale [59]. This is often the case in IoT environments, where sensors and actuators rely on knowledge that takes the form of dynamic streams of data [79]. Agents need to react in a timely fashion to these data streams, so as to update their beliefs or revise their intentions. Although certain extensions exist for allowing KGs to integrate stream reasoning [26] and continuous querying capabilities [25, 51, 9], the usage of streaming KGs by autonomous entities is still incipient. Beyond traditional pipelines of data streams, the employment of streaming KGs in multi-agent systems and in general by decentralised autonomous entities has the potential to boost the exploitation of local and personal information captured by devices on the edge and other sensing agents.

Moreover, autonomous systems need to be able to represent and reason with knowledge about external entities (the environment, other agents, organisations) but also with internal knowledge about their goals, tasks and capabilities. The former is primarily knowledge that naturally emerges from heterogeneous agents cooperating to reach some common goal. Knowledge is hence created and exchanged following agent communication and social mechanisms [16]. This knowledge may emerge from the agreements and deliberations that arise during the lifetime of autonomous entities [55]. Ontologies may then be constructed from commonly learned facts contributed by autonomous agents, thus constituting an interaction-driven KG. Given that this knowledge is built from diverse and potentially contradicting sources, different methods of reconciliation and mediation have been explored; for example by relying on ontology alignment techniques [14, 31], or distributed ontology evolution [56].

In this kind of scenario, knowledge is heterogeneous and distributed, and therefore interaction mechanisms and protocols cannot assume that agents share common knowledge beyond what is required to establish their interactions. This heterogeneity also implies that agents can use both (onto)logical or, more broadly, symbolic knowledge as well as numerical, sub-symbolic knowledge representation. The evolution of this type of knowledge cannot happen in a siloed manner, since each type of knowledge representation is essential to cover different aspects; i.e., abstracting from and making sense of stimuli from the environment. Furthermore, autonomous and cooperative decision-making implies that all the systems inhabiting the environment have to respond to the same type of demands, therefore it does not make sense for them to react to these demands in isolation, especially when they are cooperating towards a common goal.

Any knowledge evolution model should hence consider the abilities of the autonomous agents to: i) reason and evolve their own knowledge, which may be modelled using disparate formalisms; ii) reason about the different demands they perceive from the environment they are immersed in; iii) deliberate about the knowledge to evolve, given conflicting demands (and ultimately knowledge); and iv) interact with other autonomous systems to communicate and cooperatively decide the best course of action.

A recent systematic review [2] surveyed a number of knowledge-enabled approaches to develop and deploy autonomous robots in order to assess whether current state of the art ontology-based approaches model those concepts and relations that are necessary to represent knowledge about other autonomous entities and the environment; i.e., the ability to model and reason about the system’s own capabilities and the capabilities of others. This study argues that the key factors for the success of knowledge-based autonomous systems is the use of explicit ontological models to support autonomous systems in understanding and reasoning about their environment, their own internal knowledge and knowledge about other agents and organisations. The systematic review is based on Langley’s characterisation of autonomy with respect to abstract reasoning tasks such as recognition and categorisation, decision-making and choice, problem solving and planning, as well as perception and situation assessment [50]. Langley distinguishes between: i) what, i.e., the functional capabilities that an autonomous system should exhibit; and ii) the way in which these functional modules should interconnect (how). Functional capabilities and their interconnections determine the specific vocabulary needed to represent autonomy and the types of reasoning tasks that are supported by different ontological models. These interconnections are supported by the exchange of data and knowledge that is often both produced and consumed by autonomous systems.

KG schemata modelling autonomous behaviour should include the definition of entities (in terms of modules or fragments of hierarchical knowledge) that represent both static (e.g., agent, object, or scene) and dynamic (e.g., task, goal, or action) aspects of autonomy. We follow the same characterisation and we discuss the ways KGs are currently used to support autonomy, as a way to identify the limitations (which are introduced in Section 5).

As discussed in Section 2, there is no accepted definition of KG, and as a result the term KG is used to refer to different data sources, and for example, to model both real-time data generated by sensors that does not conform to some schema, or a graph of interrelated data that is modelled according to some explicit schemata that often includes standard foundational (upper-level) and domain-specific ontologies, as well as organisation-specific schemata (which might be proprietary). Foundational ontologies such as SUMO [62] or Dolce [35] and its simplified version in OWL, Dolce+DnS Ultralite⁶⁶6http://www.ontologydesignpatterns.org/ont/dul/DUL.owl are used to model domain-independent concepts, e.g., time, object or event, and support interoperability at the highest level of abstraction. Domain-specific vocabularies have emerged in many domains to provide common ground for communication and have become standards to model data for the specific domain. Examples of such standards are the Semantic Sensor Network (SSN) ontology⁷⁷7https://www.w3.org/TR/2017/REC-vocab-ssn-20171019/ [37], and the provenance ontology Prov-O⁸⁸8https://www.w3.org/TR/prov-o/ [53]. SSN is a modular ontology that also includes a self-contained core ontology modelling Sensors, Observations, Samples, and Actuators (SOSA). SSN and SOSA can support a variety of applications and use cases, including satellite imagery, large-scale scientific monitoring, and use cases across the Web of Things [43]. PROV-O is an OWL ontology that describes the classes, properties and restrictions describing the provenance data model, used to integrate provenance information in different systems and based on different contexts. More recently, the W3C WoT TD [43] standard provides the vocabulary for describing the metadata and interfaces for a “Thing”, i.e., an abstraction of an entity (physical or virtual) that interacts with and participates in the Web of Things, thereby endowing agents with higher autonomy with respect to their environment (see Section 3.1).

Regarding organisational autonomy (see Section 3.3), a similar issue manifests itself when representing and reasoning with obligations and norm governance in general [44]. The ODRL standard⁹⁹9https://www.w3.org/TR/odrl-model/ is a formal language, expressed in RDF, for modelling policies regulating the use of digital rights. However, the ODRL core that models the components of a policy expression does not provide any formal specification of its semantics, and it provides only informal semantics described in English [81]. Recently, some extensions have been proposed, based on the work in Normative MAS, where autonomy determines the decision-making processes [20]. For example, Fornara and Colombetti [33] aim to fully capture the formalisation of norms used to express obligations, permissions and prohibitions by modelling the operational semantics of conditional obligations using OWL, possibly extended with SWRL [39] rules to overcome OWL’s limited expressivity with respect to the state of an interaction and the rules for computing the state of obligations [33]. SWRL adds Horn-style rules to OWL, thus permitting the modelling of additional domain knowledge that cannot be done by OWL alone. Both OWL and SWRL adopt the Open World assumption, whereby the absence of a statement does not necessarily mean that it can be inferred to be true or false. This is in contrast with SHACL, which adopts the Closed World Assumption (a statement that is not known to be true must be false). More recently, SHACL has been proposed to model more complex type of norms, such as compensatory norms [75].

Several conceptual architectures and programming platforms that support autonomous behaviour are currently available (e.g., JaCaMo [12] or KnowRob [85]), and they rely on domain and foundational ontologies and KGs for representing and sharing declarative knowledge. However, many of the reasoning tasks underlying executive autonomy (see Section 3.4) involve procedural knowledge that is not typically captured by KGs. The execution of complex tasks involving several autonomous entities in open environments requires them to interact with each other and the environment, and to coordinate their executions; e.g., to support autonomy in delegation. Interaction protocols (also called choreographies within the service community) are formal specifications of these interactions, which describe the roles played by the autonomous systems, the allowed actions and choices, and the states and termination conditions [8]. Similarly, communication protocols specify the rules of interactions governing a dialogue between autonomous systems [29, 34, 71]. Typically, these protocols are expressed in some protocol language based on different abstractions and operational assumptions that are difficult to compare. An exhaustive description of these protocols is outside the scope of this paper, but we refer to the work by Chopra and colleagues for a review of the main formalisms and a comparison framework [21]. These protocols ground the meaning of the different concepts involved in some ontology, therefore implicitly establishing an a-priori commitment to a shared understanding of the terms being communicated [41], where the procedural part of the model and the execution traces are typically not included in KGs.

Various efforts in robotics have also attempted to model and execute procedural knowledge about some tasks, e.g., kitting, where a robot places (a set of) parts on a tray and carries them to an assembly cell, using a knowledge-enabled approach. For example, the KnowRob [85] architecture has been extended with OWL based assembly planning, where assembled products are modelled according to ontologies, which are then used to plan the next action according to faulty and missing assertions in the robot’s beliefs about an ongoing assembly task [11]. Other knowledge-enabled approaches, however, convert OWL descriptions into PDDL [58] specifications [7, 47].

Multi-dimensional event knowledge graphs (EKGs) [30] have been proposed as a way to represent process events in a way that supports the execution and inspection of process traces. Recently, EKGs have also been proposed to model process logs [45] in order to support process analysis.

KGs have also been explored as a possible source of knowledge to be injected in reinforcement learning approaches [1, 3], where autonomous agents use sequential decisions to learn and adapt through interactions with their environment by trying to find (near-) optimal policies to perform an intended task [84]. In these approaches, commonsense KGs such as Wikidata [90] or CSKG, the Commons Sense Knowledge Graph [42] are exploited to model commonsense notions such as spatial, part-whole, or temporal relations, and even affordances [19]. In line with our discussion in Section 3.5, we propose that such usage of KGs represents only the beginning of a movement to the more active curation of KGs by autonomous software agents themselves.

Recently, both KGs and architectures for autonomous agents have emerged as promising facilitators of autonomy in the context of generative AI and specifically Large Language Models (LLMs) [65, 82]. In this context, KGs are typically envisioned as providers of reliable, “hard” knowledge that LLMs are not capable of producing. While this emphasises the relevance and timeliness of research on KGs as facilitators of autonomy, we deliberately choose to present perspectives in this paper that can be considered orthogonal to LLMs and generative AI in order to avoid a tunnel view that is focused on current hype topics.