POMDP wildfire UAV benchmark with mission-aware risk-adaptive planning. Paper 2 implementation — successor to FLARE (Paper 1).
Status: In active development. See CLAUDE.md for full specification, architecture, and phase plan.
Sensor-only partial observability POMDP where a simulated UAV navigates a 2D urban wildfire scenario using modeled sensors (LDRobot LD19 LiDAR, 130° HFOV camera, BN-880 GPS), builds belief in flight, and learns an adaptive risk coefficient ρ̂(s,t) conditioned on calibrated mission physics (Arrhenius insulin decay, Weibull hemorrhage survival, CFK CO inhalation, etc.).
Resolves the Paper 1 ranking inversion under sensor uncertainty — and reveals a second one along the observability axis.
Strictly 2D simulation. No real drones, no MAVLink, no satellites, no external APIs at runtime.
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -e ".[dev,gis,ml]"
pytest tests/ -qSee SETUP section in CLAUDE.md for full instructions and verification commands.
- Now (M1 16GB): Phases 1–4 + Phase 6 (env, sensors, belief, missions, analysis, MCP server). Small RL debug runs OK on MPS.
- Later (RTX 3060 PC): Phase 5 — full MARS-RL training (10M steps × 5 seeds × 5 scenarios × 6 difficulties).
MIT