Presentation
Adaptive De-escalation Trainer: Piloting a RAG-Enhanced, Emotionally Modulated AI Simulator for Police Training
DescriptionEffective police de-escalation training requires realistic practice, often limited by traditional method’s scalability and consistency. We developed an adaptive AI simulator enabling officers to practice de-escalation techniques across varied scenarios representing common crisis encounters, such as mental health issues or suicidal ideation. Our system integrates a large language model (LLM) with retrieval-augmented generation (RAG) for contextual scenario grounding. It employs dynamic Text-to-Speech (TTS) synthesis to modulate the AI's emotional prosody. This emotional expression adapts based on keyword analysis ("success"/"failure" phrases) of officer speech and an internal state model, creating a responsive conversational partner. Our ongoing study with active police officers (target N=5-10) evaluates the system's feasibility and effectiveness. Through participant surveys (assessing interaction quality, emotional realism, workload via NASA-TLX) and system performance logs (latency, state changes, keyword impact), we investigate perceived conversational coherence/responsiveness (H1), perceived emotional realism (H2), feedback loop validation (H3), and technical viability (H4). This investigation evaluates the potential of adaptive conversational AI agents as scalable, on-demand supplementary resources for law enforcement skills training. These systems may enhance officer confidence and communication strategies by offering a secure environment for repeated practice. Concurrently, this work aims to identify limitations and areas requiring further development within this AI-based pedagogical model.
Event Type
Lecture
TimeTuesday, October 14th4:50pm - 5:10pm CDT
LocationGrand Hall L
Aerospace Systems
Cognitive Engineering & Decision Making
General Sessions
Human AI Robot Teaming (AI)