GPEvac Achieves 14.73 ms CPU Inference for Layout-Agnostic Evacuation Routing

The framework trains a single PPO policy on graph embeddings that remain permutation-invariant across building topologies. An edge-first sequential scheme propagates local crowd and threat signals before the virtual node aggregates long-range dependencies, replacing layout-specific heuristics or exhaustive search. Simulation trials across multiple floor plans recorded lower cumulative threat exposure than rule-based and standard GNN baselines.

Inference latency of 14.73 ms on commodity CPU hardware meets the timing window required for live camera feeds to update routes before the next shooter movement. The same embedding and scoring mechanism transfers without retraining to infrastructure graphs in transportation and sensor networks, confirming the architecture is not limited to evacuation.

Operational deployment hinges on integration latency with existing surveillance rather than model speed. Prior real-world systems stalled at policy rigidity; GPEvac's single-policy invariance removes that constraint. Next steps require closed-loop tests with anonymized building CCTV streams to measure exposure reduction under actual response delays.