Decentralized Fleet Coordination for Airport Ground Operations

12 ground service vehicles. No central dispatcher. Every gate covered, zero collisions. Each agent sees only its local Voronoi cell and its neighbors’ positions — the fleet self-organizes to match demand as flights arrive and depart.

The Apron

SFO international terminal modeled at scale: 7 gate clusters, 8 vehicle types (fuel, catering, deice, GPU, water, belt loader, bus, maintenance). The coverage problem — keep the right vehicles close to the right gates at the right time — is dynamic, multi-type, and decentralized by design.

Voronoi Coverage

Each agent owns the Voronoi cell of the apron closest to it. Lloyd's algorithm drives vehicles toward the weighted centroid of their cell — where the demand field is highest. As planes push back and new flights arrive, demand shifts and cells reform. No messages to a central controller; agents converge on their own.

BVC Collision Avoidance

Buffered Voronoi Cells shrink each agent's feasible region by half the distance to each neighbor — any movement within the BVC is guaranteed collision-free. Lloyd commands are projected onto the BVC via a QP. Result: the coverage law and the safety constraint compose cleanly, with no separate collision-avoidance layer needed.

Demand Response

Flight arrival triggers a demand pulse at the target gate. The demand field φ(x,t) peaks at the gate position and decays spatially. Nearby vehicles detect the shift in their centroid target and converge — no explicit task assignment, no auction protocol. The nearest available vehicle wins by physics.

Simulation Results

Zero inter-vehicle collisions across simple, medium, and stress-test scenarios. Coverage cost (H(P,φ)) converges to within 8% of the centralized optimum. Average response time to a demand pulse: 23 seconds at realistic ground vehicle speeds.