You skipped the game, truthful present is conscionable the optimal path. Is this what you would’ve done?
A fewer weeks ago, 30,954 group mowed the aforesaid lawn. This is what their paths looked like.
Many came adjacent to the optimal path. 52% of group came wrong 5 moves of the champion imaginable path, and 16% did it perfectly.
Many different approaches, yet similar, beautiful bully outcomes. Only 49 squares request to beryllium covered, but group recovered 14,589 different ways to do it. Despite each that variety, the median personification still deed 91% ratio (within 5 moves).
For this mini lawn, location are 12 different cleanable solutions and group recovered them all.
But what were you really solving?
Finding the champion way done the squares is simply a classical machine subject problem. No, don’t leave! We committedness this won’t get excessively wonky.
Path readying problems
Solving the section mowing problem is formally called Coverage Path Planning, and is related to 1 of the astir celebrated way readying problems successful machine science: The Traveling Salesman Problem.
The premise is simple: a salesman must sojourn a group of cities precisely erstwhile and return to their starting point, each while taking the shortest imaginable route. A pizza transportation driver readying their stops aliases a trick-or-treater maximizing their candy haul is solving a type of this problem.
We wanted to create this mowing research aft we read a study that recovered humans travel impressively adjacent to computer-calculated optimal solutions to this problem pinch less stops, and moreover do beautiful bully — conscionable 11% little businesslike than optimal — for 70+ stops. We recreated their floor plan below. As the stops increase, humans’ expertise decreases. Would results beryllium akin for mowing a lawn?
A flimsy detour, if we may. There are 2 main ways to tackle a problem for illustration this. Simplifying things a bit, location are algorithms which guarantee the optimal path, and heuristics which usage shortcuts to find a “good enough” path, fast.
Why do group request heuristics? If a salesperson had to sojourn conscionable 10 cities and each imaginable way has to beryllium checked, we’d amazingly request to trial 3.6 cardinal of them. That number explodes exponentially arsenic much cities are added.
15 cities and we’re astatine 1.3 trillion imaginable routes. 😱
20 cities balloons to 2.4 quintillion imaginable routes. 🤯
When problems are small, the attack doesn’t matter much, almost everything works. But arsenic complexity grows, optimal becomes intolerable and heuristics go necessary. This is existent for some humans and computers. People some autumn backmost connected good-enough strategies erstwhile cleanable is impractical.
A guided tour
Let’s replay a mowing session, director’s-cut style, pinch immoderate commentary on the way. For this walkthrough, we’ll usage a perfectly mean tally from a personification who called themselves Bones. It took them 54 moves–five much than optimal. We recorded the timestamp of each move, truthful each hesitation aliases guidance alteration near a information footprint.
We’ll rewind to the commencement and locomotion done it. Before Bones moreover made a move, they paused for 2.9 seconds. They were perfectly mean present arsenic well. Upon seeing the section revealed, they people paused to study the section and, perhaps, made a plan.
Off went Bones! There was thing to determine successful the opening agelong until the 5th square. Continue right, aliases spell down?
This is wherever paths first diverged. Bones hardly hesitated (0.7s) and decided to spell down, for illustration 20% of different people. This is simply a captious infinitesimal that we’ll travel backmost to shortly.
Bones plodded on the extracurricular separator until they reached the bottom. Their first notable region (2.4s) came erstwhile they realized they deed a dormant extremity and had to retrace their steps.
After realizing their mistake, Bones turned astir and re-mowed their first quadrate to decorativeness this section. To get to the different unfastened conception they had to backtrack done 2 much squares. Then it was mostly soft sailing arsenic they worked astir the edges, past filled successful the remaining squares successful the middle.
What differentiated the champion mowers from everyone other came down to a azygous captious decision. We tin disagreement the section into 2 main sections: near and right. Bones completed the near conception first. Because it dead-ends, the only way retired was to retrace their steps, creating an inevitable backtrack.
If we look astatine wherever each personification ended, we tin spot this play out. People 3 aliases much moves from optimal vanished scattered crossed the lawn.
While those wrong 2 moves of optimal, including the 17% who were perfect, chose to mow the correct broadside first truthful they could extremity their tally connected the left, having anticipated the dead-end trap.
One of those was a personification whose existent sanction was Sarah. We sewage successful touch, and learned that she vanished connected the near intentionally.
“I saw the dormant extremity astatine the extremity and I was like, okay, that’s wherever I request to extremity arsenic well, truthful don’t spell that measurement first.”Sarah knew a dormant extremity meant the near conception had to beryllium reserved for the extremity to debar backtracking. The region information backed her up. Her longest hesitation came a quadrate before the fork, earlier continuing to clear the correct side.
Sarah besides did thing that appeared a spot unnatural. Instead of simply continuing consecutive a fewer times, she intuitively and preemptively turned successful a snake-like pattern. This highlighted a cardinal quality betwixt real-life mowing and our simplified game, an attack that beardown mowers exploited. Sarah, an avid gamer, seemed to admit this strategy immediately.
Six lawns
That was 1 lawn. We built 5 levels, each bigger and much cluttered than the last, up to a 14x14 section pinch almost 200 squares to cover.
30,954 group made it done 2 required levels, and 8,912 group completed a fewer other prize ones. The greenish lines show the median optimality for each lawn. The ray set represents really astir group (80%) did.
We expected a decline. The Traveling Salesman Problem investigation suggests that group do worse arsenic the problem grows. But optimality held astir 90% whether the section took 26 moves aliases 177.
One caveat: we made the lawns bigger pinch different obstacle layouts astatine the aforesaid time, truthful we can’t opportunity which mattered more. Performance was really somewhat amended connected the largest section than the erstwhile two, which suggests that layout matters much than earthy size.
Our champion conjecture is the building of a section itself. Where the Traveling Salesman Problem scatters nodes randomly, a section has structure, which lets group autumn backmost connected 2 strategies: decomposition and compression.
People decomposed the section into smaller, much manageable sections and solved them incrementally. Our interviewee Sarah described this perfectly.
“I return it portion by piece...I effort to time off everything down maine finished.”On the biggest lawn, she was definitive astir why: holding the full point successful her caput wasn’t an option.
“I’m not going to beryllium capable to scheme each of this out...just moreover trying to retrieve each of it would beryllium measurement excessively challenging.”So she chunked things up. She cleared a conception and moved connected to the next. Textbook decomposition.
As group moved betwixt rounds, they whitethorn person compressed what they learned into a intelligence exemplary of the game, allowing them to place patterns successful much analyzable lawns and accommodate their gameplay.
When you think
It seemed evident to america that group who thought longer would mow better. Below, the 7,235 group who vanished each lawn, plotted by gait (seconds spent per square, which normalizes clip crossed section sizes) against really adjacent to optimal they got. Pick immoderate gait percentile connected the x-axis and the dispersed of results looks astir the aforesaid arsenic astatine immoderate other.
While location is simply a very mini inclination connected the other ends for the apical and worst mowers, if we told you really agelong personification took, you couldn’t conjecture really good they did. Technically speaking, clip unsocial explains little than 5% of the variety successful really group did (R2 of 4.9%).
We isolated conscionable extended region times (i.e., thinking) and the aforesaid inclination held, but where they paused is simply a different story. Let’s return a look astatine level 4 arsenic an example.
Top 10% — seconds paused per square
Infinity
−Infinity+
Bottom 10% — seconds paused per square
Infinity
−Infinity+
The champion mowers (90th percentile) spent clip reasoning early arsenic they traveled down the first column, past thought much astatine the juncture betwixt the 2 stone clusters connected the bottommost left: spell toward the cluttered mediate section, aliases clear the unfastened abstraction to the correct first. They were capable to intuit that they should decorativeness successful the dead-end pouch astatine the apical successful bid to minimize backtracking. As a result, they didn’t request to region erstwhile they reached the location stretch.
Meanwhile, the worst mowers (10th percentile) sped done the opening and hardly paused to deliberation until a area aliases separator forced them to, implying reacting, not planning. They only realized that the aforesaid pouch was a dormant extremity erstwhile they were already successful it, past had to stop, backtrack, and activity retired wherever to spell from there.
The champion mowers moved astatine astir the aforesaid velocity arsenic everyone else, but they revisited acold less squares. They spent their clip astatine the forks and cruised done the rest. Below are 2 group who vanished astatine the nonstop aforesaid clip pinch different results.
We heard this successful the question and reply too. Sarah, 1 of the champion mowers, spent her attraction unevenly connected purpose. She had a mentation astir why.
“The bigger spaces person much cognitive load for you to fig retired what the champion way is. Whereas the smaller spaces...there are less decisions to make.”She planned the tight, trap-prone sections cautiously and winged it connected the unfastened ones.
So it’s not astir reasoning much aliases faster, but when reasoning happens. That’s what a bully heuristic does: it spends your attraction connected the decisions that matter and thing connected the ones that don’t.
Good luck not reasoning astir this the adjacent clip you mow a lawn!
Explore
Below is simply a small playground. Pick a section and mow it, aliases research wherever group paused, backtracked, and finished.
move: 1
use the arrow keys to move
Data and Methods
Data is simply a snapshot of the unrecorded research arsenic of July 31, 2026. There was a region successful the research aft the 2nd level to cod demographic information. Later levels were optional and presented arsenic prize rounds, truthful less group played them.
The scatterplot floor plan from the Traveling Salesman study is approximated based connected the image successful the study since information weren’t provided.
Pace is seconds per required square, not per quadrate personification really mowed. If we divided by their existent move count, a longer way would some wounded their optimality and inflate their move count, which would make the comparison circular.
To find optimal efficiency, we ran algorithms that solved each section exhaustively. The optimal solution included backtracks connected immoderate levels.
The 4.8% R2 worth comes from a linear fresh of gait percentile against optimality percentile among group who vanished each lawn.
We interviewed 1 participant, Sarah, complete video aft she gave support to usage her existent name. Everyone other is referred to by the nickname they entered during the game.
You tin research the earthy personification information connected our Github.
English (US) · 
Indonesian (ID) ·