OntoBOT: Unified Ontology for Service Robotics
- OntoBOT is an OWL-based unified ontology that models and integrates tasks, actions, environments, and robot capabilities to support context-aware robot behaviors.
- It leverages description logic and established vocabularies (e.g., DOLCE, SOMA, ROS.owl) alongside SPARQL and RDF for formal and interoperable reasoning.
- Empirical evaluations across various robots demonstrate its effectiveness in dynamic task planning, affordance checking, and facilitating knowledge sharing in service robotics.
Ontology for roBOts and acTions (OntoBOT) is a unified, OWL-based ontology for personal service robotics that formally models and connects tasks, actions, environments, and robot capabilities. It was introduced to address a recurrent systems problem in domestic robotics: effective operation requires not only physical interaction, but also the ability to interpret dynamic environments, understand tasks, and choose appropriate actions based on context. OntoBOT extends existing ontologies rather than replacing them, and its stated contributions are twofold: it unifies these aspects into a cohesive ontology to support formal reasoning about task execution, and it demonstrates generalizability by evaluating competency questions across four embodied agents—TIAGo, HSR, UR3, and Stretch—showing context-aware reasoning, task-oriented execution, and knowledge sharing in service robotics (Martorana et al., 26 Sep 2025).
1. Origins, scope, and design objective
OntoBOT emerged in the context of personal service robots used in domestic settings to assist older adults and people requiring support. In that setting, hardware components such as sensors and actuators must be integrated with software systems capable of reasoning about tasks, environments, and robot capabilities. Frameworks such as the Robot Operating System (ROS) provide open-source tools that help connect low-level hardware with higher-level functionalities, yet real-world deployments remain tightly coupled to specific platforms, producing isolated and hard-coded solutions that limit interoperability, reusability, and knowledge sharing. Existing ontologies and knowledge graphs already offered structured representations of tasks, environments, and robot capabilities, and resources such as SOMA and DOLCE already provided models for activities, spatial relationships, and reasoning structures. The specific gap identified for OntoBOT was that these resources often focus on specific domains and do not fully capture the connection between environment, action, robot capabilities, and system-level integration (Martorana et al., 26 Sep 2025).
The ontology’s scope is therefore deliberately end-to-end. It aims to bridge the full pipeline of robot task execution in open, dynamic environments, going beyond isolated representations of individual components to their interplay and dependencies. In practical terms, this means that a task is not represented only as a symbolic goal, an action is not represented only as a motor primitive, and an environment is not represented only as a container of objects. Instead, OntoBOT binds procedural task structure, executable actions, environmental components, affordances, robot capabilities, and software-level execution abstractions into a single description-logic-compatible representation.
A plausible implication is that OntoBOT is best understood not as a domain ontology for one subsystem, but as an integration ontology for task execution semantics in service robotics.
2. Ontological architecture and formal representation
OntoBOT is implemented in OWL and uses description logic-based reasoning. It makes extensive use of existing vocabularies and semantic web standards, including OWL classes and properties for modeling entities and relationships, SPARQL for querying instantiated knowledge graphs in evaluation, and RDF/TTL for data representations. Alignment with external ontologies is achieved via owl:equivalentClass, owl:subClassOf, or direct reuse of URIs. The design favors reuse over invention: minimal new elements—3 classes and 7 properties—are introduced, with the ontology instead reusing components from DOLCE, SOMA, ROS.owl, and PKO/PPlan (Martorana et al., 26 Sep 2025).
Its principal classes and relations are explicitly defined. obot:Agent is a subclass of dul:Agent and prov:Agent, denoting a robot or acting entity. obot:Environment is a subclass of dul:Place, modeling the physical space containing components. obot:Component represents physical interactable entities such as objects or furniture. obot:Affordance is a subclass of soma:Affordance. obot:Task is aligned with prov:Activity and is implemented as a procedure with sequential pplan:Steps. At the relation level, obot:actsOn links an action to a component, obot:requiresAffordance links an action to an affordance, obot:enablesAffordance links a capability to an affordance, and obot:hasNode links an agent to a ROS node. Ordering properties such as pko:nextStep and obot:nextAction structure procedures and action sequences.
This architecture is explicitly multi-level. OntoBOT represents high-level, goal-oriented activities as tasks decomposed into executable procedural steps; atomic operations as actions; the physical context as environments and components; and platform-specific functional abilities as capabilities. The ontology thus encodes both declarative and procedural structure without collapsing one into the other. It also models robot software integration by representing each agent with its ROS nodes and communications, linking high-level affordances and low-level ROS-executable capabilities. This indicates that OntoBOT complements middleware-level abstractions by giving them a formally queryable semantic layer.
3. Affordances, capability grounding, and execution semantics
The ontology’s unification principle is centered on affordances. Tasks are represented as procedures involving steps and actions; each action targets an environment component and requires certain affordances; and an action is executable only if the environment affords it and the robot can physically perform it through a matching capability. This is the core execution semantics of OntoBOT: feasibility is jointly determined by task structure, environmental affordances, and embodied capability (Martorana et al., 26 Sep 2025).
The canonical textual example is “retrieve spoon from drawer.” In this representation, the task decomposes into actions such as opening the drawer and grasping the spoon. These actions require openable and graspable affordances, which depend both on the properties of the drawer or spoon and on the robot’s abilities. OntoBOT thereby supports inference of which actions are feasible in a given environment with a given robot, enables stepwise checking of which parts of a task can or cannot be executed, and supports partial task assignment, dynamic replanning, and interpretability when execution fails.
This affordance-centered organization places OntoBOT within a longer line of ontology-based robotics research. Earlier affordance-based ontology work encoded functional substitution through shared affordances, for example by representing umbrella, raincoat, hat, and old newspaper via hasFunction: RainwaterProtector, so that a robot could select a non-substantive substitute when the substantive object was unavailable (Hidayat et al., 2012). Separately, component-based capability ontologies infer capabilities from owned components and low-level capabilities, using patterns such as transitive hasComponent, component–capability equivalence, and direct tie-in to affordance relationships; those patterns were explicitly described as generalizable, including for adoption in OntoBOT (Dussard et al., 2023).
This suggests that OntoBOT’s novelty lies less in introducing affordances or capabilities as isolated notions than in coupling them to procedural task knowledge, environmental modeling, and ROS-level execution abstractions within one aligned ontology.
4. Empirical evaluation and competency-question analysis
The reported evaluation instantiated knowledge graphs for two domestic scenarios—“prepare breakfast” and “reorganise the kitchen”—and queried them with six competency questions. These questions asked: what objects and their affordances are involved in “prepare breakfast”; what is the required action sequence for the activity; what capabilities are necessary for the activity; which robot or robots can execute all required actions; whether a given robot can execute both activities given its capabilities; and, if not, which capabilities are missing and which steps cannot be performed. SPARQL queries were authored for each competency question and run over the instantiated knowledge graph (Martorana et al., 26 Sep 2025).
The evaluation covered four platforms with varying physical capabilities:
| Robot | Capabilities | Limitation in evaluation |
|---|---|---|
| TIAGo | Grasp, hold, place, open/close, pour | None for the two tested activities |
| HSR | Grasp, hold, place, open/close | Does not pour |
| UR3 | Grasp, hold, place, pour | Does not open/close |
| Stretch | Grasp, hold, place | Does not open/close or pour |
The results were explicit. Only TIAGo had the capability set to do all steps in both activities. HSR lacked pouring and therefore could not perform “Serve food.” UR3 could not open or close but could serve food. Stretch lacked some required affordances, preventing complete execution. The results were reported as consistent with the modeled affordance-capability relationships and with real-world system limits.
The generalizability claim rests on reuse across heterogeneous robots rather than on retraining or remodelling the environment for each platform. The same task, activity, and environmental models could be reused, with only robot capabilities needing update. In encyclopedic terms, this is a strong form of representational portability: the ontology is intended to hold the environment and task semantics fixed while varying embodiment-specific capability assertions.
5. Knowledge-graph use, multimodal integration, and interoperability
Later work used OntoBOT as the formal schema in a neurosymbolic framework that combined multimodal LLMs with ontologies and knowledge graphs for platform-independent robot reasoning. In that setting, OntoBOT encompassed objects, appliances, furniture, and components; properties such as dul:hasComponent, obot:hasAffordance, geo:sfContains, obot:onTopOf, and dul:hasLocation; and action/workflow constructs such as dul:Action, obot:Workflow, and obot:Instruction, with precedence relations obot:precedes and obot:follows (Martorana et al., 13 Jul 2025).
That framework generated two graph types: observation graphs, representing the perceived current environmental state, and action graphs, representing ontology-grounded sequences of atomic actions needed to complete a task. Four neural-symbolic interaction strategies were evaluated: Dynamic Path Extractor (DPE), Description to KG (D2KG), D2KG with retrieval-augmented generation (D2KG-RAG), and Image to KG (I2KG). Structural validation used RDF validity, triple count, ontology compliance, ontology coverage, and SHACL shapes for both observation and action graphs. GPT-o1 and LLaMA 4 Maverick consistently outperformed other models, and the reported findings also indicated that newer models do not guarantee better results, highlighting the critical role of the integration strategy in generating ontology-compliant knowledge graphs.
These later results clarify an important point about OntoBOT’s position in the stack. Symbolic systems such as ontologies and knowledge graphs support structured, standardized, and sharable representations, but they struggle with raw and noisy sensory input. Multimodal models are better suited for interpreting images and natural language, but often lack transparency, consistency, and knowledge grounding. OntoBOT functions here as the ontology-compliant target representation that makes multimodal perception outputs reusable across systems. The implication is interoperability in a strong semantic sense: any robot or software that understands OntoBOT can interpret and execute the resulting plans, subject to its own capabilities.
6. Relation to adjacent ontology-based robotics frameworks
OntoBOT belongs to a broader field of ontology-driven robotics in which several adjacent problems have been studied with different emphases. Linked Planning Domain Definition Language (PDDLS) connects actions, predicates, and objects to global IRIs and uses SPARQL and SHACL to establish conditional predicates and action-state compatibilities in robotic planning (Tatsubori et al., 2019). Temporal ASP approaches integrate action theories with ontological knowledge in EL^\bot, allowing non-deterministic actions, causal rules for ramifications, and polynomial encodings that guarantee action consistency with respect to an ontology (Giordano et al., 2021). Generic human-robot collaboration ontologies extend robot-centric models with explicit human branches, hazardous and safe situations, safeguards, and recovery protocols for dynamic task and workflow management (Skarzynski et al., 2020). A survey of action representations in robotics identifies a persistent lack of standardized formal representations and recommends structured, ontology-like action models as a route toward effect-centric, grounded, and generalizable action knowledge (Zech et al., 2018).
Against that background, OntoBOT is specific in its emphasis. It is not principally a safety ontology, a temporal answer-set formalism, or a planning language extension. Its distinctive target is the connection between environment, action, robot capabilities, and system-level integration in personal service robotics. This suggests one common misconception to avoid: OntoBOT does not claim to solve perception, control, safety assurance, or formal action verification on its own. Rather, it provides a unified representational substrate that can be combined with perception pipelines, reasoning engines, workflow systems, and robot software architectures.
A second misconception is that generalizability here means unrestricted transfer. The reported evidence concerns four embodied agents, two domestic scenarios, and six competency questions. A plausible implication is that broader deployment in less-controlled homes, multi-robot settings, or safety-critical human-robot collaboration would require extensions of the kind explored in adjacent work rather than a simple scaling of the original evaluation. Within its stated scope, however, OntoBOT defines a reusable ontology pattern for context-aware reasoning, task-oriented execution, and knowledge sharing in service robotics (Martorana et al., 26 Sep 2025).
