Section 47.4: Coverage and inspection; multi-drone coordination

"Covering an area is mowing a lawn from the air: the hard part is missing nothing and counting it honestly."

A Careful Control Loop

Ask four drones to photograph every square metre of a collapsed parking structure and the failure that ends the mission is rarely a crash: it is the thin unscanned stripe that nobody notices until the mission log already reads "coverage 100 percent." Coverage, inspection, and multi-drone coordination live or die on that gap between claimed and achieved, so this section treats them as a concrete embodied AI skill rather than a label. The core contract is: observe coverage grid, camera footprint, communication state, and battery reserve, allocate viewpoints and paths across agents, and judge the result with coverage completeness and duplicate-inspection rate.

Check the earlier frame, control, and model chapters against the exact interface used here: state variables, timing budget, action limits, and evaluation panel.

Figure 47.4.1 turns coverage and fleet coordination into inspectable artifacts: area decomposition, assignment policy, communication state, deconfliction rule, battery budget, and coverage certificate.

Theory

Coverage planning asks for a path that sweeps a camera footprint over every point of a target region. The classic exact method is boustrophedon decomposition (the word means "as the ox plows", back and forth). The free region is split into cells with no obstacles inside, and each cell is covered by parallel sweep lines, called boustrophedon, or lawn-mower, passes, spaced by the sensor footprint width \(w\). The drone flies one lane to the far edge, steps over by \(w\), reverses, and repeats. For a convex cell of width \(W\) the number of lanes is \(\lceil W / w \rceil\), and the turns happen only at the cell boundary, which keeps the path efficient.

This approach has been deployed in practice. Skydio's commercial bridge and infrastructure inspection platform uses a variant of viewpoint-coverage planning in which the drone autonomously identifies structural faces, generates required camera poses, and sequences a path through them while maintaining a safe standoff distance. The system reports coverage as a percentage of required poses achieved, mapping directly to \(\eta\) below. DJI agriculture spraying drones use the same boustrophedon structure but widen lane spacing to reduce flight time, accepting a small uncovered fraction and relying on overlapping spray cones to compensate. Both systems illustrate the same core trade: completeness versus battery endurance.

The success criterion is the coverage fraction: the area actually seen divided by the area that should be seen,

$$\eta = \frac{\lvert \text{covered cells} \rvert}{\lvert \text{target cells} \rvert}.$$

A complete plan reaches \(\eta = 1\). Lane spacing controls the trade between \(\eta\) and flight time: spacing equal to the footprint gives full coverage with no overlap in the ideal case, while spacing wider than the footprint leaves uncovered stripes (\(\eta < 1\)) and tighter spacing wastes battery on redundant passes. Real footprints depend on altitude and camera field of view, so coverage planning and altitude planning are coupled; flying higher widens \(w\) and reduces lanes but lowers ground resolution. To make the stakes concrete: a single drone covering a 100-metre square field with a 2-metre footprint needs 50 lanes and roughly 18 minutes of flight, but adding a second drone and splitting the Voronoi partition cuts that to 9 minutes per drone, while a third drone brings it to 6 minutes. Fleet size, not algorithm cleverness, is the dominant lever on total mission time.

Coverage completeness \(\eta\) is only half of the evidence panel promised for this skill; the other half is the duplicate-inspection rate \(\rho\), defined as the fraction of covered cells seen by more than one drone: \(\rho = \lvert \text{cells seen} \geq 2 \text{ times} \rvert / \lvert \text{covered cells} \rvert\). A fleet can report \(\eta = 1\) while \(\rho\) is high, meaning several drones redundantly re-swept the same ground while a different stripe went unseen by anyone; a fleet with the same \(\eta\) and low \(\rho\) genuinely divided the work. Both numbers belong in the same logged panel, because \(\eta\) alone cannot distinguish an efficient partition from a wasteful one.

So which lever actually matters most when a deadline is closing in: a smarter algorithm or launching one more drone? The numbers above suggest fleet size dominates, but before reading on, consider what happens if that extra drone arrives mid-mission with no shared coverage map.

That question only has a precise answer once the fleet has a rule for dividing territory among drones and reassigning it as conditions change; the next paragraph introduces exactly that rule, the Voronoi partition and auction handoff.

A drone that covers its assigned cell perfectly while its neighbor covers the same cell twice has not solved coordination: it has just made redundancy expensive and gaps invisible.

Multi-drone coordination adds an assignment layer on top of single-drone coverage. The standard approach partitions the target region into Voronoi cells centred on each drone's current position. Each drone owns the sub-region closest to it, so the partition adapts as drones move. Within its cell, each drone runs an independent boustrophedon sweep. When a drone's battery falls below a return threshold, it broadcasts a handoff request. Neighbours with enough reserve then bid on the uncovered remainder using an auction protocol: each bidder submits an estimated completion cost, and the lowest-cost bidder wins the cell. Separation is the critical safety constraint during reassignment. Two drones must not enter the same cell at once, so the communication channel enforces a short exclusion window. Most research systems build this same three-layer structure: Voronoi partition, per-cell sweep, and auction-based handoff. The multi-UAV coverage work by Scherer et al. (2015) and the distributed coverage framework in Schwager et al. (2011) both follow it.

The auction protocol matters in embodied AI because the fleet must make the handoff decision in real time under communication latency, battery pressure, and imperfect position estimates. A naive round-robin assignment ignores travel distance, so the "next" drone might sit on the opposite side of the field. It then burns its entire reserve flying to the abandoned cell and cannot complete it. In a representative 500-metre field trial, round-robin handoff typically left around a third of the area unfinished because the assigned drone ran out of battery mid-transit, while a single auction round reduced the uncovered area to a few percent by always picking the nearest eligible bidder; the exact figures depend on fleet size, field geometry, and wind, so treat them as illustrative rather than a universal bound. A centralized planner avoids this but creates a single point of failure: one dropped radio packet stalls the whole fleet. The auction is a middle ground that is both distributed and cost-aware, so the fleet degrades gracefully when agents drop out rather than freezing.

Mechanically, it is a single-round sealed-bid process. When drone \(i\) broadcasts a handoff request, each eligible neighbour \(j\) computes a bid \(c_{ij}\) from travel time to the cell centroid plus estimated sweep time for the remaining lanes, and transmits it once. Drone \(i\) picks the minimum, acknowledges, and departs for base. The winner \(j^*\) waits out the exclusion window before entering, so only one agent occupies the cell at a time.

Because each handoff reshapes who owns which cell, the partition the auction feeds into cannot be the fixed map it is often imagined to be.

Worked Example

Simulate a boustrophedon sweep over a 10x10 grid and measure the coverage fraction. Lane spacing equal to one cell (every row swept) should give full coverage; doubling the spacing should leave half the area unseen. Running both makes the spacing-versus-coverage trade concrete.

# Boustrophedon (lawn-mower) coverage over a 10x10 grid.
def boustrophedon(N, lane_step):
    covered = [[False] * N for _ in range(N)]
    path = []
    for row in range(0, N, lane_step):          # one lane per swept row
        cols = range(N) if (row // lane_step) % 2 == 0 else range(N - 1, -1, -1)
        for col in cols:                        # serpentine direction per lane
            path.append((row, col))
            covered[row][col] = True
    seen = sum(cell for r in covered for cell in r)
    return path, seen / (N * N)

N = 10
path, eta = boustrophedon(N, lane_step=1)        # spacing = footprint
print(f"lanes touch every row : path length {len(path)}, coverage = {eta:.0%}")
print(f"first turns           : {path[8:13]}")

_, eta_wide = boustrophedon(N, lane_step=2)       # spacing = 2x footprint
print(f"spacing doubled       : coverage = {eta_wide:.0%}  (stripes missed)")

Expected output: a full-coverage path at unit spacing and a degraded coverage fraction when the lanes are spread out. The coverage fraction is the evidence field: report it alongside flight time, because a plan that finishes faster by widening lanes is only a win if \(\eta\) still meets the inspection requirement.

Practical Recipe

1. Write the skill contract: observable variables, action interface, metric, allowed recovery actions, and stop conditions.
2. Build the smallest baseline that can fail in an interpretable way.
3. Run the maintained library version with the same inputs, scenarios, and metric code.
4. Add one perturbation aimed at the expected failure: two drones optimize local coverage and create unsafe convergence near the same structure.
5. Save one artifact containing config, seeds, logs, summary metrics, and two representative traces.

Coverage and inspection; multi-drone coordination becomes robust when the chapter separates three claims. The conceptual claim explains why the skill should work. The systems claim explains which interface changes. The evidence claim records which same-panel metric would convince a skeptical builder.

Keep flight physics, airspace constraint, battery state, timing, wind, and safety monitor inside the evidence artifact rather than in a post-run explanation.

When Coverage and inspection; multi-drone coordination fails, do not collapse the whole method into one score. Assign the failure to a subsystem, rerun one perturbation that isolates the suspected cause, and keep the trace as a reusable diagnostic case.