Hierarchical Physical Common Sense Ontology

• Hierarchical physical common sense ontology is a formal taxonomy that organizes physical entities, events, and properties in a recursive, multi-level hierarchy.
• It enables robust commonsense inference and diagnosis in AI systems by integrating rigorous frameworks like DOLCE, Cosmos-Reason1, and space–motion models.
• The ontology is extensible via domain-specific instantiations, optimizing physical reasoning in applications such as robotics, language understanding, and embodied AI.

A hierarchical physical common sense ontology is a systematic, formalized taxonomy for representing and reasoning about the entities, properties, and dynamics of the everyday physical world, supporting robust commonsense inference for both human-centric AI systems and cognitive modeling. Such ontologies aim to encode the structurally recursive, multi-level organization of space, time, objecthood, material constitution, eventhood, and physical law, enabling the diagnosis, evaluation, and augmentation of physical reasoning abilities in both knowledge bases and learned models. Foundational versions—such as DOLCE and Cosmos-Reason1—provide a precise (often axiomatized) backbone, while domain instantiations extend or operationalize the hierarchy for tasks in robotics, language understanding, and embodied AI (Borgo et al., 2023, NVIDIA et al., 18 Mar 2025, Suchan et al., 2017, 0808.1211).

1. Foundational Principles and Ontological Commitment

Hierarchical physical common sense ontologies embody strong commitments to cognitive plausibility, interoperability, and formal rigor. The classical top-level division, exemplified by DOLCE (Borgo et al., 2023), partitions reality into several mutually exclusive, collectively exhaustive categories:

• Endurants (Continuants): Entities wholly present at any time, subdivided into independent (e.g., physical objects and amounts of matter) and dependent (qualities, features) subtypes.
• Perdurants (Occurrents): Entities extended in time; events and processes.
• Abstracts: Non-physical, such as quality spaces, sets, and propositions.

The central tenets are:

• Disjointness and exhaustivity (each entity fits exactly one root branch),
• Parent–child subsumption for identity and inheritance,
• Explicit meta-properties (e.g., rigidity, dependence) to enforce correct generalization,
• Recursively extensible to admit fine-grained domain-specific subtrees while preserving foundational coherence (Borgo et al., 2023).

2. Structural Organization and Core Hierarchies

The organization of physical common sense is typically expressed as a directed acyclic graph or tree with clearly defined levels. Canonical examples:

DOLCE (Borgo et al., 2023)

Level	Category	Description/Examples
Top	Endurant	Objects, matter, features, qualities
	Perdurant	Processes, events, states
	Abstract	Quality spaces, sets, propositions
Sub-level	Physical Endurant	Physical Objects, Amounts of Matter
	Physical Perdurant	Process, State, Achievement, Accomplishment
	Physical Quality	Mass, color, shape (inhering in endurants)

Cosmos-Reason1 (NVIDIA et al., 18 Mar 2025)

Category	Subcategories	Coverage Examples
Space	Relationship, Plausibility, Affordance, Environment	Spatial layout, object support
Time	Actions, Order, Causality, Camera, Planning	Temporal sequence, causal inference
Fundamental Physics	Attributes, States, ObjectPermanence, Mechanics, Electromagnetism, Thermodynamics, AntiPhysics	Object status, physical law verification

All subcategories correspond to specific capabilities (e.g., reasoning about mechanical stability or event order), tightly coupled to evaluation and data-curation schemas in physical AI (NVIDIA et al., 18 Mar 2025).

Space–Motion Ontology (Suchan et al., 2017)

This framework for human-robot interaction implements a layered entity hierarchy (spatial, temporal, space-time histories, body models), linked to a rich inventory of spatial, mereotopological, temporal, and motion primitives, supporting declarative reasoning and learning from sensor data.

Commonsense Linguistic Ontology (0808.1211)

Entities are partitioned into Abstract (Event, State, Process, Activity) and Physical (Natural, Artifact) branches, with explicit sort declarations for subsumption and disjointness, as well as type annotation for all predicates and relations, enabling grounded compositional semantics.

3. Formalization: Axioms, Constraints, and Reasoning

Hierarchical physical ontologies are rigorously axiomatized to capture identity, composition, parthood, participation, and quality assignment.

Sample Axioms (Borgo et al., 2023, Suchan et al., 2017, 0808.1211)

• Temporary Parthood:

$$\forall x, y, t.\; PartAt(x, y, t) \rightarrow Endurant(x) \land Endurant(y) \land Time(t)$$

• Proper Part:

$$ProperPartAt(x, y, t) \leftrightarrow PartAt(x, y, t) \land \neg PartAt(y, x, t)$$

• Participation in Events:

$$\forall x, y, t.\; ParticipatesIn(x, y, t) \rightarrow Endurant(x) \land Perdurant(y) \land Time(t)$$

• Qualities:

$$\forall q, x.\; PhysicalQuality(q) \rightarrow \exists!x (q \, inheresIn \, x)$$

• Disjointness:

$Endurant \cap Perdurant \subseteq \perp$

Space–motion ontologies further specify relational algebra for RCC-5/8 (e.g., disconnected, external contact), Allen interval algebra for temporal relations, OPRA for orientation, and event schemas for interaction chains, formalized in higher-order logic and CLP(QS) (Suchan et al., 2017, 0808.1211).

Type-theoretic linguistic ontologies use sort subsumption and rule-based type unification, integrating compositional semantics via rules asserting relation saliency and existence inference (0808.1211).

4. Illustrative Classification and Instantiation

Concrete instantiation is achieved by classifying physical and eventive entities into the hierarchy via subsumption, axiomatizing their properties, and formalizing interactions.

Examples (Borgo et al., 2023, NVIDIA et al., 18 Mar 2025)

Physical Entity	Ontological Path
Car	Car ⊑ PhysicalObject ⊑ IndependentEndurant ⊑ Endurant
Wood	Wood ⊑ Matter ⊑ IndependentEndurant
Driving Event	Driving ⊑ Process ⊑ Stative ⊑ Perdurant
Stability Judgment	Space:Plausibility capability (Cosmos-Reason1)
Object Permanence	FundamentalPhysics:ObjectPermanence (Cosmos-Reason1)
Reach_For Action	holds_in(reach_for(P, O), Δ) schema (Space–Motion)

Such categorizations enable robust query, inference, and learning mechanisms: from grounding sensory input in space–motion axioms (Suchan et al., 2017), to tracing model benchmark failures to individual ontology leaves (NVIDIA et al., 18 Mar 2025).

5. Methods for Specialization and Extension

The extensibility of foundational ontologies is realized through modular addition of domain-specific subtypes, qualities, and events while maintaining logical coherence:

• Subsumption: Extend classes (e.g., RobotArm ⊑ PhysicalObject; WeldingProcess ⊑ Process) (Borgo et al., 2023).
• Property Addition: Introduce new qualities (e.g., GripForceQuality ⊑ PhysicalQuality) and augment quality spaces.
• Domain Events: Attach bespoke process/event types relevant for verticals (e.g., CrackPropagation for materials science).
• Axiom Patterns: State new parthood, participation, disjointness, and unity conditions in line with parent classes.

Preservation of interoperability follows from adherence to meta-properties such as rigidity and dependence, and documentation and modularization best practices, frequently operationalized in semantic web frameworks (e.g., OWL:imports) (Borgo et al., 2023). Cosmos-Reason1’s closed-world physical capabilities ontology enables strategic data coverage and targeted curriculum learning for physical AI (NVIDIA et al., 18 Mar 2025).

6. Applications in Embodied Reasoning and Language

Hierarchical physical common sense ontologies underpin a range of AI capabilities:

• Benchmarks and Evaluation: Cosmos-Reason1 establishes diagnosis-oriented VQA and action prediction benchmarks explicitly grounded in the subcategories of its ontology, structurally decomposing “physical reasoning” scores into interpretable sub-capabilities (NVIDIA et al., 18 Mar 2025).
• Robotics and HRI: Deep Semantic Abstractions ontology enables declarative reasoning and relational learning in robotics and human-object interaction, supporting grounding of sensor data into spatio-temporal event schemas via logical inference and abduction (Suchan et al., 2017).
• Natural Language Understanding: Commonsense physical ontologies support compositional, type-unified semantics, directly handling phenomena such as metonymy and copredication by revealing hidden ontological relations (0808.1211).
• Interoperability/Integration: DOLCE underpins standards and public-domain resources (e.g., CIDOC CRM, DBpedia, WordNet), emphasizing its centrality to broad-based data/knowledge integration (Borgo et al., 2023).

7. Impact, Current Practices, and Prospects

Hierarchical physical common sense ontologies provide the formal scaffolding essential for explainable, compositional, and verifiable AI reasoning about the physical world. Their explicit structure enables fine-grained diagnosis and targeting of model deficiencies (e.g., strong on Mechanics, weak on Thermodynamics (NVIDIA et al., 18 Mar 2025)), supports industrial and research interoperability (Borgo et al., 2023), and underlies active lines of research in embodied AI, natural language semantics, and robotics (Suchan et al., 2017, 0808.1211).

Recent methodological trends include benchmarking training and evaluation directly against ontology leaves, closed-world category enumeration for exhaustive coverage, and the integration of logical axiomatization with data-driven approaches. The durability and transferability of foundational ontologies such as DOLCE, along with operational frameworks like Cosmos-Reason1, suggest their continued centrality to progress in commonsense physical reasoning.