Capability-Based Taxonomy Overview

• Capability-Based Taxonomy is a structured classification that hierarchically organizes skills and dispositions in agents and systems, supporting modular representation and interoperable reasoning.
• It underpins applications in robotics, AI, and industrial systems by decomposing processes into tasks, skills, and primitives, thereby facilitating dynamic task allocation and safety evaluation.
• The taxonomy formalizes safety zoning and cognitive frameworks, guiding risk assessment, domain-specific performance benchmarking, and governance across diverse application areas.

A capability-based taxonomy provides a systematic classification and hierarchical structuring of capabilities—basic powers, skills, or dispositions—in agents (human, artificial, or hybrid), artifacts, and systems. Such taxonomies formalize the concept of "capability", support modular representation, enable structured reasoning over agent skills, and create a principled foundation for interoperability, benchmarking, and safety assurance across diverse application domains.

1. Foundational Concepts and Ontological Structuring

The foundational basis for capability-based taxonomy is the ontological analysis of "capability" as a particular subclass of disposition—a tendency, potential, or power realized in specific processes. The distinction between a disposition and a capability is that the latter is a disposition whose realization is of interest to some organism or stakeholder. In the hierarchy of the Basic Formal Ontology (BFO), capability sits as a subclass of Disposition, with further subsumption by Function (for etiologically grounded capabilities), as follows:

• Disposition
  • Capability: $$Capability(x) \iff Disposition(x) \ \wedge\ \exists o,p \bigl(Organism(o) \wedge Process(p) \wedge Realization(p,x) \wedge hasInterest(o,p)\bigr)$$.
  • Function: Capabilities that exist due to a bearer’s evolutionary or design history.
  • BiologicalFunction: Functions of biological entities.
  • ArtifactFunction: Functions of artifacts reflecting intentional design.
  • NonFunctionCapability: Capabilities with no etiological functional grounding.
  • BiologicalCapability, ArtifactCapability, CollectiveCapability, BodyguardCapability.

The ability to structurally distinguish between functions, non-functional but interest-relevant capabilities, and unused ("bodyguard") capabilities enables cross-domain reasoning and supports domain-specific extensions, ensuring ontological and practical interoperability (Beverley et al., 2024).

2. Capability Taxonomies in Robotics and Industrial Systems

Robotics and industrial automation communities have converged on multi-level capability taxonomies to standardize description, integration, and execution logic for agent skills. The canonical hierarchical arrangement is as follows (Pantano et al., 2022, Dussard et al., 2023):

• Process: Abstract, solution-neutral workflow consisting of Skills (e.g., "gearbox assembly").
• Task: Resource-specific and parameterized instance of a Process, with concrete robot mapping.
• SkillGroup: Clusters of similar Skills for descriptive reference only.
• Skill: Named, parameterizable robot capability (e.g., Pick, Place, MeasurePose).
• ParameterizedSkill: Instantiation of a Skill with specific parameters (object, pose, speed).
• Primitive: Hardware-level atomic operations (OpenGripper, MoveLinear).

This structuring enables:

• Modular decomposition (Process $\rightarrow$ Task $\rightarrow$ ParameterizedSkill $\rightarrow$ Primitive)
• Explicit parameterization (differentiating fixed vs. parametric Skills)
• Integration with safety and industrial requirements (e.g., force-limited Primitives)
• High-level semantic linkage using Description Logic (DL)/OWL: Capabilities are inferred from the presence of required components and sub-capabilities, allowing fully automated reasoning about what an agent (robot, system) is capable of, given current resources and attached objects (Dussard et al., 2023).

This approach underpins automatic task allocation, adaptive affordance inference, and robust safety evaluation in industrial scenarios. Empirical survey work demonstrates that the "placement-pick" family (Pick & Place) dominates current research, but increasing industry demand for high-mix, low-volume flexibility is accelerating development of truly parametric and context-sensitive skill taxonomies (Pantano et al., 2022).

3. Cognitive and Multimodal Taxonomies in AI Systems

Capability-recursive taxonomies for AI, particularly in LLM agents or multimodal in-context learning (ICL), are structured around both neurocognitive and compositional principles.

The six-level capability-oriented taxonomy developed for unified multimodal ICL spans:

1. Perception: Explicit perceptual anchoring (e.g., visual grounding).
2. Imitation: Schema or style replication (e.g., generating captions in a given style).
3. Conception: Fast mapping of new symbols to previously unseen entities.
4. Deduction: Rule extraction and application through multi-step sequences.
5. Analogy: Latent transformation abstraction and application to novel instances.
6. Discernment: Human-aligned subjective evaluation given by demonstrations with chain-of-thought explanations.

Formally, distinct LaTeX criteria quantify each level; for example, perception is tied to attention mass over annotation regions, imitation via style-embedding consensus, conception by detection of correct new token application, and so on (Xu et al., 25 Mar 2026).

This taxonomy provides not only a conceptual lens but drives dataset and benchmark construction (e.g., UniICL-760K, UniICL-Bench) and enables principled architectural interventions (e.g., the Context-Adaptive Prototype Modulator) to stabilize cross-modal ICL performance. Empirical studies demonstrate that this structure ensures more explainable, robust, and scalable few-shot learning (Xu et al., 25 Mar 2026).

4. Capability-Based Zoning for Safety and Governance

To address high-impact and hazardous capabilities in AI systems, a zoning taxonomy operationalizes risk-aware information exchange and governance (Pistillo et al., 2024). The taxonomy decomposes capability progress into four graduated "zones", where each capability is emergently located by its proximity to realized harm:

Zone Name	z(c) Range	Key Thresholds / Example
Safe Zone	[0, 0.25)	No multi-step planning; GPT-2
Precursory Zone	[0.25, 0.50)	Planning/self-modeling; maze
High-Risk Zone	[0.50, 0.75)	Deception; exploit chaining
Red Line	[0.75, 1.00]	Real-world exploit/prohibition

A formal scoring function $z(c)$ combines performance, resource use, and autonomy to produce a continuous measure. Policy and security actions, information flow, and regulatory responses are strictly mapped to each zone, creating transparent early warning and mitigation pathways. This zoning aligns with risk protocols in critical safety industries and is being implemented in national/international AI Safety Institutes (AISIs) for staged, cross-developer capability sharing and red teaming (Pistillo et al., 2024).

5. Domain-Specific Capability Frameworks (Healthcare, Education, etc.)

In domain-adapted contexts such as empirical evaluation of LLM-based agents in healthcare, multidimensional capability taxonomies extend the foundational ontology to explicit operational and empirical axes. The seven-dimensional taxonomy introduced by Vatsal et al. (Vatsal et al., 4 Feb 2026) decomposes agentic AI competencies into:

1. Cognitive Capabilities (planning, perception, action, meta-monitoring, conflict resolution)
2. Knowledge Management (external integration, memory, updating)
3. Interaction Patterns (dialogue, event triggers, human-loop, error recovery)
4. Adaptation & Learning (drift, RL, few-shot)
5. Safety & Ethics (guardrails, bias, privacy, compliance)
6. Framework Typology (multi-agent, central orchestration)
7. Core Tasks/Subtasks (documentation, QA, diagnosis, treatment, monitoring, benchmarking)

Each sub-dimension is empirically scored (Fully/Partially/Not Implemented) across 49 evaluated systems, surfacing prevalence statistics, capability co-occurrence, and performance asymmetries, e.g., robust external knowledge integration (76% fully implemented) vs. deficient event-triggered activation (92% not implemented) (Vatsal et al., 4 Feb 2026).

Similarly, in cognitive assessment taxonomies for human-AI learning scenarios, each cognitive level (e.g., Bloom's Taxonomy) is split into individually realizable and distributed (human+AI) realizable variants, with an additional orchestration meta-level. This "Augmented Cognition Framework" makes explicit the asymmetric dependency whereby distributed (AI-augmented) competence typically presupposes individual competence, and incorporates diagnostic rubrics for learning outcome assessment (Ayodele et al., 31 Jan 2026).

6. Inference, Linkage, and Compositional Patterns in Capability Taxonomies

All major frameworks define linkage rules between capabilities, sub-capabilities, components, and external objects/entities:

• Formal DL/OWL axioms in robotics relate $$CapabilityType \equiv \exists hasComponent.C_1 \ \sqcap \ldots$$ and propagate capability possession up via property chains such as $$hasAvailableComponent \equiv isCapabilityOf^{-1} \circ hasComponent$$ (Dussard et al., 2023).
• Low-level vs. high-level capability composition is systematic: low-level skills are component-driven, high-level skills require sub-capabilities and additional components.
• Affordance generation integrates agent side (capability possession) with environment side (object dispositions) using SWRL rules; e.g., a grasp affordance arises if the agent has a MoveObjectViaGraspingCapability and the object is Pickable.
• Dynamic recomputation: If a robot carries a new external component (e.g., picks up a tool), its capability set is automatically re-inferred, updating possible affordances and available tasks (Dussard et al., 2023).

This compositional logic, supported by reasoners and standard rule engines, underlies on-the-fly task assignment, safety rationale updates, and scaling to complex collaborative human-robot or distributed cognitive scenarios.

7. Implications, Significance, and Forward Outlook

Capability-based taxonomies establish a rigorous, interoperable core for modeling, evaluating, and governing agent abilities. Core implications across domains include:

• Unified structuring and inference across modalities (robotics, AI, cognition, governance).
• Automated composition and reasoning, essential for self-awareness, dynamic planning, and safety in intelligent systems.
• Measurable, empirical benchmarking of deployed systems via standardized sub-dimensions and explicit scoring rubrics.
• Transparent governance pathways, especially in high-impact and safety-critical settings, with actionable information flow tied to continuous capability assessment.

Ongoing research is prioritizing modular, extensible capability models that support both domain-neutral ontological reasoning (Beverley et al., 2024) and practical, benchmark-driven instantiations in target domains (e.g., medicine, education, safety auditing), advancing toward robust trust calibration, orchestration, and transfer learning diagnostics in human-AI symbiosis (Ayodele et al., 31 Jan 2026, Xu et al., 25 Mar 2026).

---

References:

• (Dussard et al., 2023) Ontological Component-based Description of Robot Capabilities
• (Pantano et al., 2022) Capability-based Frameworks for Industrial Robot Skills: a Survey
• (Beverley et al., 2024) Capabilities: An Ontology
• (Xu et al., 25 Mar 2026) UniICL: Systematizing Unified Multimodal In-context Learning through a Capability-Oriented Taxonomy
• (Pistillo et al., 2024) Pre-Deployment Information Sharing: A Zoning Taxonomy for Precursory Capabilities
• (Vatsal et al., 4 Feb 2026) Agentic AI in Healthcare & Medicine: A Seven-Dimensional Taxonomy for Empirical Evaluation of LLM-based Agents
• (Ayodele et al., 31 Jan 2026) Revising Bloom's Taxonomy for Dual-Mode Cognition in Human-AI Systems: The Augmented Cognition Framework