Aero-Grid

An applied drone delivery routing engine combining evolutionary search, graph pathfinding, supervised classification, and reinforcement learning inside an interactive web application built with FastAPI, Next.js 15, React 19, and HTML5 Canvas. It plans multi-stop drone delivery missions across a 40×40 urban grid with static buildings and restricted no-fly zones.

Flight viability is evaluated by three supervised classifiers — Gaussian Naive Bayes, Logistic Regression, and Decision Tree — trained on 17,520 historical weather records from the Open-Meteo Archive for Islamabad, predicting flight safety across three classes: Safe to Fly, Requires Altitude Drop, and Grounded. Delivery order itself is optimized with a Genetic Algorithm using order crossover, tournament selection, and elitism, cutting flight distance by 50.9% against a naive sequential tour.

Per-leg navigation runs on A* search across three selectable heuristics — Octile, Manhattan, and Euclidean — to visually demonstrate admissibility constraints and node-exploration trade-offs under 8-directional movement. A tabular Q-Learning agent with potential-based reward shaping is trained from scratch per leg as a comparison against A*, reaching within 7.4% of optimal path length and surfacing where a learned policy generalizes well versus where it fails under dynamic environment changes.