Physical AI capabilities can emerge from vastly different formation histories, yet taxonomies based solely on morphology, architecture, learning algorithm, task or domain fail to pinpoint the underlying source. We define a capability‑formation source as a factor that materially contributes to the emergence of a capability, distinct from components or construction steps.
Using reconstructive induction until theoretical saturation, we identified seven non‑exclusive sources: Recorded‑Experience (RE), Predictive‑Modeling (PM), Evaluative‑Interaction (EI), Surrogate‑Environment (SE), Mechanism‑Grounded (MG), Embodied‑Coupling (EC) and Evolution‑Driven (ED).
After literature search, deduplication and coding rule establishment, we performed three rounds of maximum‑difference and negative‑case sampling, covering challenges such as curriculum and self‑supervised learning, active inference, open‑ended and developmental learning, planning and search, neuro‑symbolic architectures, digital twins, generative physical‑world models and morphology‑control co‑design. Within the scope fixed on September 4 2026, all 49 evidence records were explainable by the seven sources individually or in combination; no irreducible eighth source emerged and no new core definition or boundary rule was required, indicating theoretical saturation (though not logical completeness or exhaustive future coverage).
The framework separates similarity of observed capability from similarity of formation process, providing a tool for analysis of explanation, transfer, replication, dependencies, governance evidence and geoeconomic foundations.
Review: This taxonomy offers researchers a clear lens to trace the roots of physical AI abilities, facilitating cross‑domain transfer and systematic evaluation.