EmSim: AI-Driven VR Simulation for De-Escalation Training in Law Enforcement

De-escalation is a critical communication skill for law enforcement, yet traditional training methods are resourceintensive and offer limited feedback. We investigate the use of AIdriven virtual reality (VR) to provide scalable, adaptive training for high-stakes encounters. We developed EmSim, a scenariobased VR system where trainees engage in open-ended dialogue with an emotionally responsive virtual subject. A rubric-guided evaluator powered by a large language model assesses each utterance and modulates scenario escalation. In a mixed-methods study with 14 trainees and 5 trainers, we analyzed usability, workload, behavioral trajectories, and post-session reflections. Results revealed two user strategies, careful de-escalation and boundary testing, and showed how trainees adapted across trials. Findings highlight how AI-assisted XR can support reflective, feedback-rich communication training.