UrbanFlood OpenEnv: RL Flood Crisis Management Simulator
Deterministic city-scale hydrodynamic crisis simulator for multi-agent reinforcement learning policies.
¶ 01Context & Design Rationale
“Emergency logistics systems fail not from lack of vehicles, but from the cascade of uncoordinated local decisions during rapid inundation. Simulating evacuation requires deterministic state propagation at scale.”
UrbanFlood OpenEnv was architected to replace sluggish continuous-fluid CFD models with a discrete graph topology that retains topological hydraulic fidelity while executing at over 100,000 environment steps per second. By vectorizing the state transition matrix into unified GPU tensors, policy gradient algorithms (PPO / SAC) converge in hours instead of days.
¶ 02Architectural Pipeline
SCALE: CITY-WIDE CELLULAR GRAPHTOPOLOGY INGEST
OSM Vector Road Network
Graph Extraction
→
HYDRO-GRID
2D Inundation Tensor
GPU Vectorized
→
AGENT DISPATCH
Rescue Fleet & Evacuees
Dec-POMDP Policy
→
GYM WRAPPER
Observation & Reward Vector
Step Latency < 10µs