Efficient Angle-Aware Coverage Control for Large-Scale Map Reconstruction using Drone Networks

Reconstructing a field in 3D from the air needs images of every point taken from several viewing angles — one angle reconstructs badly, several reconstruct well. Angle-aware coverage control makes that a control problem: the drones fly a five-dimensional virtual field of viewpoints, three spatial dimensions plus the horizontal and vertical angle a point is seen from, and every cell of it carries an importance index that decays as cameras cover it.

The angle-aware coverage problem: a drone at position p_i observing a target field B from horizontal angle theta_h and vertical angle theta_v — The problem. A drone flying at fixed altitude observes points in the target field B from a horizontal angle θ_h and a vertical angle θ_v — the two extra dimensions that make the coverage problem five-dimensional.

The problem this paper fixes

The existing controller has a blind spot on large fields. Because the objective function never evaluates how far away an unobserved point is, distant points carry no more urgency than nearby ones. Small unobserved patches end up scattered across the field, and the drone spends its remaining flight time crossing the field to pick them off one at a time — wasteful in both time and battery. On a small testbed this barely shows. On a large field it dominates.

An obvious fix — writing distance directly into the objective — was tried and failed: the drones stopped exploring and stuck to fixed positions.

The fix

Rather than change the objective, the paper changes the nominal input of the QP. Classical Voronoi coverage supplies a move-to-centroid term that pulls each drone toward the weighted centroid of its own Voronoi cell, so distance enters through where the drone wants to go while the angle-aware constraint on the decay rate of the objective still governs what it is allowed to do.

Control architecture: a central computer updates importance indices, and each drone runs Voronoi-based coverage control feeding a QP-based controller — A central computer updates the importance indices, since they depend on every drone's history. Everything after that — the Voronoi nominal input and the QP solve — runs on each drone.

Results

Colour maps of the importance index over time for the previous controller and the proposed controller — Importance index over time, previous controller (left of each pair) against this one (right). The previous controller leaves scattered unobserved patches; this one sweeps the field in an orderly front.

In ROS simulation the difference is consistent:

Scenario	Previous controller	This controller
Single drone, large field	J ≈ 0 at ~1300 s	~900 s
Three drones, large field	~470 s	~330 s
Single drone, physical testbed	~180 s	~150 s

The multi-drone case matters because it answers the obvious objection — that simply adding drones would shrink each drone’s territory enough to hide the problem. It does shrink it, and the advantage holds anyway.

The Tokyo Tech Robot Zoo Sky testbed and the schematic of the experimental system — The hardware: Parrot Bebop 2 drones in the Tokyo Tech Robot Zoo Sky testbed, OptiTrack motion capture at 120 fps, laptops as distributed computation nodes because the Bebop's onboard chip accepts only basic velocity input.

The testbed is only 2.2 m × 2.2 m, so the field was made effectively larger by dropping the drones to 30% of the simulated altitude — narrowing each field of view enough to preserve the ratio between what a drone sees and the size of the field. The gap narrows at that scale, as expected, but it does not close.

Faster completion is also a battery argument, which is the constraint that actually limits these missions in the field.