A study investigated the formation of capabilities relevant to Physical AI. The research defined a capability-formation source as a factor materially contributing to capability formation, separate from components or construction steps. Seven sources were identified: Recorded-Experience, Predictive-Modeling, Evaluative-Interaction, Surrogate-Environment, Mechanism-Grounded, Embodied-Coupling, and Evolution-Driven. The research utilized reconstructive induction and coding rules to analyze 49 evidence records. Challenges considered included curriculum and self-supervised learning, active inference, and planning and search.
Source: https://arxiv.org/abs/2609.09627