Send a drone into a room and ask it to look at everything: every corner, every wall, every patch of floor. That deceptively simple request — cover the whole space — sits at the heart of indoor scanning and search. There are two very different ways to solve it, and at claudeDrone we have flown both: a classical, hand-written algorithm that has worked for decades, and a policy that learns by trial and error. Neither is universally better. This is an honest account of where each one shines.
Mowing the lawn
Picture mowing a lawn. You push the mower in a straight line until you reach the fence, step over one stripe’s width, turn around, and push back the other way. Stripe by stripe, you eventually touch every blade of grass. That is exactly the lawnmower coverage algorithm, known formally as a boustrophedon traversal — Greek for “as the ox turns while ploughing.” Walk a row, hit a wall, drop down, reverse direction, repeat.
The clever part is what happens when a pillar or a piece of furniture blocks the stripe. The algorithm doesn’t give up; it runs a short BFS (breadth-first search) to route around the obstacle to the next unvisited point on its sweep, then carries on. Simple, deterministic, easy to reason about.
How well does it work? On a known map, astonishingly well. In our dev-log 05 we built a compact lawnmower evaluator — only about fifty lines of code — and ran it across five maps and several start positions. Given a generous step budget it reached roughly 98.7% coverage, and it was remarkably consistent: the spread between its best and worst map was tiny. When you already know the geometry of the space, the lawnmower is close to the best you can do — the perfect yardstick, the reference any smarter method has to beat.
So why bother learning?
If a fifty-line classical algorithm gets to 98.7%, why pour effort into reinforcement learning at all?
The honest answer from dev-log 05 is sobering: early on, our learned policy lost to the lawnmower at every step budget — trailing by a wide margin on the longer runs. For a while we had been measuring our progress against our own earlier attempts rather than a strong external reference, and the gap was larger than we assumed.
But that comparison hides the catch that matters. The lawnmower cheats, in a sense: its sweep is computed directly from the map’s coordinates, so it effectively knows the whole floor plan in advance. Real indoor missions rarely offer that luxury. Rooms are irregular, layouts are unknown until you fly them, furniture moves, and the same drone may be dropped into a space it has never seen.
That is precisely the situation reinforcement learning is built for. Our learned agent sees only what its onboard sensors report — local distance readings and a memory of where it has been — and has to discover the room’s structure as it goes. Trained across many randomized layouts, the policy generalizes to new spaces rather than memorizing one. A hand-coded sweep cannot adapt to a floor plan it was never told about; a learned policy can.
Learning the shortcut
The turning point came when we looked at why the lawnmower was winning. Its strength is the long, uninterrupted pass — it commits to a straight line and rides it all the way to the wall. Our early agent, by contrast, nudged forward one cell at a time, which on a large room cannot reach far enough within a tight step budget. That was a limitation of how moves were defined, not of learning itself.
So in dev-log 07 we gave the agent a new move: go forward until something blocks you. One command could now carry it across many cells at once, mirroring the lawnmower’s signature long stroke. The effect was dramatic. On the short budget the learned policy jumped to about 56% coverage — roughly 2.15× the lawnmower’s score at the same budget — and on the medium budget it pulled clearly ahead too. On the very longest budget the classical sweep still held a small lead, but the picture had flipped on the budgets that matter most for fast indoor scans.
The most satisfying detail: in the evaluation footage the agent flew a snake-like pattern — long straight passes, a turn, then a fresh pass — the boustrophedon traversal it was never explicitly taught. Trained with PPO (Proximal Policy Optimization) across many random maps, it rediscovered the classical sweep on its own.
Which should you choose?
Neither approach is a free win. A quick guide:
- Known, open, static map? The classical lawnmower. It is simple, transparent, and close to optimal — little reason to train a network to do what fifty lines already do well.
- Unknown or irregular indoor layout? The learned policy. It adapts to floor plans it has never seen and still produces the efficient long passes that make sweeping effective.
- Dynamic scenes — moving obstacles, targets, changing conditions? Reinforcement learning, clearly. A fixed sweep cannot react; a trained policy can.
- Tight scan budget on a large indoor space? The learned agent with its long-forward move now leads on the short and medium budgets.
- Need a benchmark first? Always build the lawnmower. It is the cheap, honest reference that tells you whether your fancy method is earning its complexity.
These methods are partners, not rivals. The classical heuristic gave us a number to chase, and chasing it forced us to find the move that finally let learning surpass it. Build the simple thing first — then let it tell you whether the complicated thing is worth it.
Further reading: dev-log 05 — the lawnmower baseline and dev-log 07 — the multi-cell forward move.