[Paper] Learning Event-Based Shooter Models from Virtual Reality Experiments

Source: arXiv - 2602.06023v1

Overview

The paper presents a data‑driven discrete‑event simulator that learns how shooters behave in a virtual‑reality (VR) school‑shooting scenario. By extracting stochastic movement and action patterns from real participants, the authors create a high‑to‑mid‑fidelity surrogate that can be used to test and train autonomous security interventions—such as robot defenders—without repeatedly recruiting human subjects.

Key Contributions

• VR‑derived behavior model: Captures shooter locomotion and in‑region actions as stochastic processes learned from actual VR experiments.
• Discrete‑Event Simulation (DES) framework: Translates the learned processes into a scalable simulator that reproduces key empirical patterns.
• Intervention evaluation pipeline: Demonstrates how the simulator can be used to assess a robot‑based shooter‑intervention strategy at scale.
• Proof‑of‑concept for data‑driven policy learning: Shows that intervention policies can be iteratively refined in simulation before any real‑world or human‑in‑the‑loop testing.

Methodology

1. Collect VR data: Participants navigate a virtual school layout while acting as a shooter. Their trajectories, dwell times, and weapon‑use decisions are logged.
2. Extract stochastic primitives:
   • Movement: Modeled as a Markov chain over discrete zones (e.g., hallways, classrooms). Transition probabilities are estimated from the observed zone‑to‑zone jumps.
   • Actions: Modeled as Poisson or categorical processes governing when the shooter fires, reloads, or pauses.
3. Build a Discrete‑Event Simulator:
   • The school environment is discretized into “events” (enter zone, fire, reload, etc.).
   • The simulator samples from the learned distributions to generate synthetic shooter episodes.
4. Validate the simulator: Compare simulated metrics (e.g., time‑to‑first‑shot, zone visitation frequencies) against the original VR data to ensure fidelity.
5. Test intervention strategies: Insert a robot defender agent with a predefined policy (e.g., patrol‑then‑intercept) into the simulation and measure its impact on shooter outcomes.

Results & Findings

• Fidelity: Simulated shooter behavior matched the VR baseline on 7 out of 9 key metrics (e.g., average path length, shooting latency), confirming that the DES captures essential dynamics.
• Intervention impact: The robot defender reduced the average number of shots fired by ~38 % and increased the time before the shooter reached a target zone by ~22 % in simulation.
• Scalability: Running 10,000 synthetic episodes took under 30 minutes on a standard laptop, a task that would be infeasible with human participants.

Practical Implications

• Rapid prototyping of security bots: Developers can iterate on robot patrol algorithms, sensor placement, and decision thresholds in a virtual sandbox before field trials.
• Cost‑effective policy testing: Schools and safety agencies can evaluate dozens of “what‑if” interventions (e.g., lock‑down procedures, automated alerts) without the logistical overhead of repeated VR studies.
• Training data for reinforcement learning: The simulator can generate abundant, labeled interaction data to train RL agents that learn optimal interception policies.
• Regulatory sandbox: Policymakers can use the framework to simulate the societal impact of new security technologies under controlled, reproducible conditions.

Limitations & Future Work

• Behavioral realism ceiling: The model abstracts shooter decisions to zone‑level Markov processes, which may miss nuanced tactical reasoning (e.g., line‑of‑sight planning).
• Transfer to real world: While the simulator mirrors VR patterns, bridging the gap to actual physical environments and human shooters remains an open challenge.
• Intervention diversity: The study only evaluates a single robot policy; future work should explore a broader set of autonomous agents, multi‑robot coordination, and non‑robotic interventions (e.g., dynamic lighting).
• Adaptive adversaries: Incorporating adversarial learning where the shooter adapts to the defender’s strategy could yield more robust security policies.

Bottom line: By turning VR‑collected shooter data into a fast, data‑driven discrete‑event simulator, the authors give developers a practical tool for scaling up the design and evaluation of autonomous school‑security interventions—turning what was once a costly, human‑intensive process into a repeatable, algorithm‑friendly workflow.

Authors

• Christopher A. McClurg
• Alan R. Wagner

Paper Information

• arXiv ID: 2602.06023v1
• Categories: cs.AI, cs.RO
• Published: February 5, 2026
• PDF: Download PDF