Autonomous Emergency Response & Search/Rescue

The Problem & Inspiration

High-fidelity search planning relies on data, such as Koester’s Lost Person Behavior and Coast Guard drift models. Currently, using these requires manual reference tables or specialized GIS software. AERO//SAR automates this workflow, putting these physics and statistical models behind a natural language interface so dispatchers can act without being GIS experts.

Core Workflow & Architecture

1. Natural Language Dispatch The dispatcher inputs the incident in plain English (e.g., "28-year-old male hiker, last seen at Angels Landing trailhead, 5 hours overdue"). LLM parses the narrative to execute a 5-tier geocoding search and establishes the Initial State $ x_0 $ as a delta function centered on the Last Known Point (LKP)

2. Terrestrial Dynamics: The Markovian Cost Surface For land searches, we model movement as a Discrete Time Markov Chain (DTMC). The state vector $ x_t $ evolves across a variable-resolution grid (default ~30m):

$$x_{t+1} = x_t \cdot T$$

The transition matrix $ T $ is weighted by Tobler's Hiking Function rather than simple linear distance. Velocity decreases as slope increases:

$$V = c \cdot \exp(\alpha \cdot |\tan(\theta) + \beta|)$$

This base velocity is then modified by subject profile parameters (e.g., a "despondent" subject is less deterred by steep terrain than a "dementia" patient).

3. Maritime Dynamics: Lagrangian Particle Tracking For water incidents, we replace the grid with a stochastic simulation of 2000 particles. Each particle's velocity $ \mathbf{V}_{obj} $ is calculated using OSCAR ocean currents and CCMP wind data, adjusted by a Leeway Factor $L$:

$${V}_{obj}^{(k)}(t) = {V}_{current}(x_k, t) + L \cdot {V}_{wind}(x_k, t) + \epsilon_k$$

Euler integration (via OpenDrift) updates positions over time to account for drift and turbulence $ \mathbf{\epsilon}_k $.

4. Autonomous Drone Dispatch The system converts the resulting Probability Density Function (PDF) into actionable search zones using Isosurface Extraction.

ALPHA: Executes a "Hasty Search" ($ R_{0.75} $ zone) along the primary trail.
BRAVO: Sweeps the high-probability "Hot Spot" ($ R_{0.95} $ zone).
CHARLIE: Establishes "Wide-Area Containment" ($ R_{0.99} $ zone).

Challenges & Lessons

Land Calibration: Tuning the seven interacting movement parameters for different subject categories required calibration.
Data Wrangling: Merging real-time ocean NetCDF files and ensuring geocoding resolved to trailheads rather than mountain summits was a hurdle.
Key Finding: Markov chains are ideal for fast, noise-free probability modeling over terrain, while particle methods are superior for the inherently stochastic nature of maritime drift.