Mapping the Knowledge Structure of Physical Artificial Intelligence: A Data-Driven Systematic Review

Physical artificial intelligence (PAI) has emerged as a transformative paradigm that integrates AI into physical entities, enabling direct interactions with real-world environments. However, despite rapid expansion across diverse domains, PAI research has remained highly fragmented and failed to provide a comprehensive understanding of its overarching knowledge structure. To address this gap, this study conducted a data-driven systematic review of 317 publications indexed in the Web of Science between September 2020 and October 2025. For the analysis of annual publication volume, the growth trend was assessed using complete calendar-year observations from 2021 to 2024, while the 2025 publication count was reported separately as a partial-year observation through October. Combining bibliometric network analysis with latent Dirichlet allocation topic modeling, we identified seven latent research topics. We integrated these fragmented topics into a unified, three-layered hierarchical architecture encompassing (1) physical interaction and infrastructure, (2) policy learning and control, and (3) cognitive integration and multimodal reasoning. The temporal analysis revealed a distinct evolutionary trajectory, indicating a structural shift from simulation-based, navigation-centric studies toward greater cognitive and multimodal integration and the practical implementation of embodied physical systems. This study provides a quantitative and structural mapping of PAI, offering a foundational framework to inform future interdisciplinary research and technological convergence.