DIGITAL TWIN TRAINING AND LEGAL GOVERNANCE FOR HUMANOID POLICE ROBOTS IN PATROL AND ARREST-SUPPORT

Humanoid police robots are emerging as Physical AI platforms because they can operate in human-centered infrastructure such as stairs, doors, corridors, vehicles, and control panels. This study is important because police robots do not merely perform technical tasks; they may mediate state authority, personal data processing, scene preservation, and possible restrictions on bodily liberty. The goal of this paper is to design a digital twin-based training and governance framework for humanoid police robots in patrol and arrest-support operations in South Korea. The central hypothesis is that patrol functions can be trained toward relatively high autonomy only if the digital twin encodes legal permission states, privacy requirements, evidence preservation, and human-command boundaries together with locomotion and perception; conversely, arrest judgment, bodily restraint, and hazardous device activation should not be automated under the current Korean legal framework. The study uses an interdisciplinary design-science method combining a review of digital twin and humanoid learning research, doctrinal analysis of Korean constitutional, criminal procedure, police, privacy, AI, and compensation law, and synthesis of a deployment-oriented architecture. The main findings are a five-layer digital twin, an autonomy-permission matrix, a benchmark-and-phase-gate validation matrix, and a human-in-command operating procedure. Practically, the framework helps police agencies, vendors, and regulators convert broad safety and legality requirements into testable design controls before public pilots. The paper concludes that the most defensible near-term model is not an autonomous arrest robot but a digitally trained police-support humanoid with law-in-the-loop constraints, privacy-by-design controls, and audit-ready evidence logs.