MapGyver

AI-Powered Lost Person Modelling

MapGyver uses AI to predict where a lost person is, with 73% accuracy in the first 6 hours of a disappearance. It combines terrain analysis (elevation, vegetation density), real-time weather data and behavioural psychology from the ISRID database into probability heat maps for allocating SAR resources. It was developed in response to unresolved missing persons cases in the Tasmanian wilderness.

Celine never came back.

Celine Cremer, 31, was a Belgian backpacker who’d spent a year exploring Australia. On June 17, 2023, she parked her car at Philosopher Falls in remote northwestern Tasmania and set off on what should have been a short bushwalk. She was never seen again.

Accuracy at 6hr
73%
vs 45% traditional
Cases tested
47
Resolved SAR
Search area
1,000km²
After 6 hours
Phone found
60m
From last signal

Nine days passed before anyone raised the alarm. Her family only realised something was wrong when she failed to board her scheduled Spirit of Tasmania ferry. Police found her car still sitting in the trailhead car park.

What followed was one of Tasmania’s most intensive search operations: hundreds of volunteers, Tasmania Police with specialist teams, drones, helicopters with thermal imaging, and a cadaver dog flown in from New South Wales. They searched the dense bushland around Philosopher Falls for weeks and found nothing.

No trace of her was found.
Dense forest canopy
Dense forest canopy — GPS signal degradation zone — Tasmania wilderness
Terrain reconnaissance — Field footage
Navigation challenges — Field footage

Inspector Andrew Hanson told media that phone data suggested Cremer had strayed from the main walking track, possibly using a navigation app to take a shortcut back to her car as daylight faded.

1h2h3h4h5h6hlast known point
The radius grows with the hours, the area with their square.

Wilderness search and rescue works against the clock. Every hour that passes, the possible search area quadruples. A person can walk roughly 3 kilometres per hour in moderate terrain, so after six hours the area is over 1,000 square kilometres of dense bush.

No team has the resources to cover that, so searches prioritise, and much of the prioritising is guesswork. The traditional approach treats the lost person as a static point: draw circles, send teams, expand outward. But Celine was not static. She was moving, making decisions and following terrain she did not understand.

In December 2025, more than two years after Celine vanished, a renewed search by family and friends found something. Volunteer Tony Hage spotted her purple Samsung phone, approximately 60 metres from her last known coordinates.

Through all the helicopters, search teams and technology, her phone had been sitting less than the length of a football field from where she was known to have been. The dense Tasmanian bush had hidden it completely.

Police theory now suggests Celine dropped her phone and continued without it, becoming disorientated in terrain that all looks the same to someone panicking. The case has been reopened. The open question is whether she could have been found faster.

last known pointcreek
Tracks wander, double back, and mostly run downhill to water.

An expanding circle assumes two things about lost person movement: that it is random, and that it is uniform in direction. Neither holds.

Decades of search and rescue data, compiled in databases like ISRID, show patterns that hold across thousands of cases. Disorientated people follow rules, even when they do not know they are following them.

Downhill travel
78%
Within 3km/12hr
85%
Water seeking
72%
Path following
65%

ISRID Database, Koester et al.

Seventy-eight percent of lost hikers travel downhill. It often leads deeper into trouble, but to a panicking person going downhill feels like progress.

Eighty-five percent are found within three kilometres of their last known point after twelve hours. Disorientation creates loops: people who think they are walking in a straight line are circling.

These patterns are statistical certainties. In Celine’s case, nobody was using them.

The model predicts where a person will go rather than bounding where they could have reached. Reachability grows with the square of elapsed time and says nothing about direction; behaviour is directional and the statistics above say which way.

Each agent runs a persona as a system prompt (a profile of experience, condition and pressure) and sees only what a person on that ground would see: nearby slope, the sound of water, the light going.

Experienced hiker
“You’ve been hiking for 20 years but took a wrong turn 3 hours ago. Stay calm, follow water downstream. Mark trees with your knife. You have 6 hours of daylight, half a water bottle, and two energy bars.”
Missing child
“You are 8 years old. You wandered from the campsite chasing a butterfly. Everything looks the same. You’re scared and crying. Adults always said ‘stay put if lost’ but you hear water and you’re thirsty.”
Solo backpacker
“You’re using a navigation app to shortcut back to your car. Daylight is fading. The terrain is steeper than it looked on screen. You drop your phone. Do you go back for it, or push on?”

An experienced hiker makes calculated decisions and a child follows impulse. A solo backpacker with a dying phone makes choices that look irrational until the pressure on them is taken into account.

Agents are placed on real terrain at a last known position and run until they stop or the light goes.

Vision is restricted to line of sight over the heightfield. Energy, hydration and daylight decrement per move, and each decision is taken against the state those leave the agent in.

Ten thousand runs produce ten thousand tracks. Their density is the output: a probability surface over the search area, concentrated wherever terrain repeatedly funnels agents to the same ground.

Overhead — terrain and track
Ground level — what the agent can see

Overhead, the track wanders and doubles back; at ground level the same terrain is a wall of near-identical slope. Aggregate ten thousand of these and the probability concentrates toward the drainage corridor below the last known position.

The model was validated against 47 resolved cases. It was given the last known position, terrain, weather and a persona, but not the outcome. A case counted as a hit when the recovery point fell in the top-ranked band of the predicted surface.

Prediction accuracy by search window

WindowTraditionalMapGyver
6 hours45%73%
12 hours38%67%
24 hours31%58%

At six hours the model put the recovery point in the top band in 73% of cases against 45% for expanding circles. The margin narrows with elapsed time as the behavioural signal decays: 67% against 38% at twelve hours, 58% against 31% at twenty-four.

At six hours the difference is 28 points of hit rate. Search teams are resource-bound, so a hit rate is a number of teams pointed at ground worth walking.

Celine Cremer is still missing.

Her phone was found 60 metres from where she last had signal, on ground that search teams walked past for two years. If MapGyver had been deployed in the first week, with the terrain modelled, the behavioural patterns applied and the probability maps generated, would the teams have looked in the right place?

That cannot be answered. The case is one of the 47 only in the sense that it motivated the work; it is unresolved, so it cannot be scored.

MapGyver is not deployed and has not been used in an operation. It has a validated hit rate against resolved cases and a method a search controller could act on. Field use needs a partner agency, and the failure mode is worth naming before anyone relies on one: a confident surface over the wrong drainage.

This research is conducted in collaboration with search and rescue professionals. For partnership inquiries: research@drksci.com