Industries / Government, education and community / Government / Emergency services
Search and rescue
Finding people who are lost, trapped or in distress, at sea, in mountains and wild country, underground and in collapsed buildings, and getting them out.
7
Milestones achieved, from 1824 to 2022.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1824
A charity for lifeboats
Sir William Hillary's plan for a national service to save lives at sea became a charity at the London Tavern in Bishopsgate, London, on 4 March 1824. It is now the Royal National Lifeboat Institution (RNLI), whose lifeboats around Britain and Ireland are crewed by volunteers. Local lifeboats already existed: the Zetland, built in 1802, is the oldest surviving lifeboat in the world, by the RNLI's account. By its own count the RNLI has saved more than 146,000 lives at sea since 1824.
Credit
Sir William Hillary; the RNLI
1991
Rescue teams agree common rules
Urban search and rescue (USAR) teams find and free people trapped in collapsed buildings. After specialized teams from several countries worked side by side in the earthquakes in Mexico in 1985 and Armenia in 1988, the International Search and Rescue Advisory Group (INSARAG) was set up in 1991, with its secretariat in the UN Office for the Coordination of Humanitarian Affairs (OCHA) in Geneva. UN General Assembly resolution 57/150 of 16 December 2002 endorsed its guidelines. Since 2005 international teams have been classified Light, Medium or Heavy through a voluntary peer review, starting with Hungary's team that year. INSARAG's list showed 64 classified teams on 18 November 2026: 34 Heavy, 24 Medium and 6 Light.
Credit
INSARAG; UN OCHA
1994
Eighteen minutes under snow
Markus Falk, Hermann Brugger and L. Adler-Kastner published an avalanche survival curve in Nature in 1994: the chance of surviving as a function of how long a person stays completely buried. As summarized in the Canadian Medical Association Journal (CMAJ) in 2011, survival stays above 91% for the first 18 minutes, then falls to 34% between 19 and 35 minutes as most buried people suffocate. Current guidance on rescuing and resuscitating buried people rests on these Swiss data, the CMAJ authors note. Their comparison of 301 Canadian and 946 Swiss burials from 1980 to 2005 found that survival in Canada began to fall earlier, at about 10 minutes, which they linked to more injuries and the denser snow of coastal mountains, and they called for digging people out within 10 minutes.
LessonA rescue time limit measured in one country's snow can be too generous in another's.
Credit
M. Falk, H. Brugger and L. Adler-Kastner; Canadian comparison by P. Haegeli and colleagues
2008
Where lost people go
Robert Koester's book Lost Person Behavior (2008) sorts lost people into 41 categories, such as hikers, hunters, children and people with dementia, and gives for each the distances from the point last seen within which a quarter, half and three quarters of them were found. It rests on the International Search & Rescue Incident Database (ISRID), which held more than 50,000 cases when Koester compiled it in the 2000s with US Department of Agriculture funding. A researcher told Undark, a science magazine, in 2024 that in the search and rescue research community it is "mandatory to have as a source". Koester's October 2025 count puts ISRID at 532,816 incidents, 231,398 of them from the US Coast Guard.
LessonAverages pooled from many places can mislead in terrain unlike the average, as research on Yosemite National Park has argued.
Credit
Robert Koester, dbS Productions
2009
Satellites stop hearing 121.5 MHz
Older distress beacons sent an analog signal on 121.5 MHz, and the satellites of Cospas-Sarsat, the international satellite system that picks up distress beacons, were built to hear it. Those beacons were "a source of chronic false alerts from the outset", says the US National Oceanic and Atmospheric Administration (NOAA), so in 2009 Cospas-Sarsat stopped listening for them and kept the digital 406 MHz beacons, whose alerts can usually be checked against the owner's registration. Even so, NOAA says 98% of 406 MHz beacon activations are false alarms, most of them settled with a phone call to the registered owner.
LessonA distress signal helps only if rescuers can quickly check whether it is real, which is why beacons are registered.
Credit
International Cospas-Sarsat Programme; NOAA, which runs the US part of the system
2021
Robots search underground courses
The DARPA Subterranean Challenge, run by the US defense research agency, set robot teams to explore tunnels, subway-like structures and natural caves with no GPS and poor radio links, find objects DARPA had placed there and report where they were. In the final on 24 September 2021, Team CERBERUS, run by a single human supervisor with four ANYmal legged robots and flying robots, found and located 23 of the 40 objects. CSIRO Data61 also scored 23 but reported its last object a little later and lost the tiebreak; MARBLE scored 18.
LessonEven the winning robots missed 17 of the 40 targets.
Credit
DARPA, which set and scored the course; Team CERBERUS, led by Kostas Alexis, from the University of Nevada, Reno, ETH Zurich, the Norwegian University of Science and Technology, UC Berkeley, the University of Oxford, Flyability and Sierra Nevada Corporation
2022
Satellite SOS from a phone
At about 2 a.m. on 1 December 2022, Alaska State Troopers were told that a man traveling by snowmachine from Noorvik to Kotzebue had become stranded and had sent an alert with Emergency SOS via satellite on his iPhone. Apple's Emergency Response Center passed his position to the Northwest Arctic Borough search and rescue coordinator, four volunteer searchers found him, and no injuries were reported. The trade site Wireless Estimator called it the first reported use of the feature, then available on the iPhone 14 in North America.
Credit
Apple; Alaska State Troopers; Northwest Arctic Borough search and rescue volunteers
Watch for
Second-generation distress beacons
Cospas-Sarsat says that from 2 November 2026 its system is at initial operational capability for second-generation beacons, which use MEOSAR, the search and rescue receivers carried on the Galileo, GPS, GLONASS and BeiDou navigation satellites. It says the new beacons send more information and give a more accurate position. MEOSAR alerts have been in operational use since December 2016, and the program expects to declare MEOSAR fully operational as early as 2026.
Achieved whenapproved second-generation beacons are on sale and a national rescue authority publishes the location accuracy and alert times seen in real activations.
Claimed by
International Cospas-Sarsat Programme, November 2026
AI sorts drone photos
In July 2025 the mountain and cave rescue service of Piemonte, Italy (CNSAS) resumed the search for a climber missing on Monviso since September 2024. Two drones took more than 2,600 photos over 183 hectares in five hours, software flagged dozens of spots whose color or texture stood out the same day, and a red dot in one flagged photo turned out to be the climber's helmet, which led rescuers to his body. A drone pilot told the BBC the software "can even hallucinate some things", and Croatia's mountain rescue service said such programs give too many misleading results in its karst terrain.
Achieved whena study across many searches reports how many flagged spots were real and whether searches with the software found people sooner than searches without it.
Claimed by
CNSAS Piemonte, July 2025, reported by BBC Future in January 2026
Robots that find casualties
The DARPA Triage Challenge, a three-year prize competition begun in 2022, has teams use sensors on drones and ground robots to find, locate and assess injured people after a mass casualty event. Its systems final ran at Disaster City, a training site near College Station, Texas, in November 2026, with scenarios from a nighttime war zone to a neighborhood hit by a tornado. DARPA gave top marks in the 2025 systems event to the DART team.
Achieved whenDARPA publishes the final scores, and a rescue service tests one of the systems in an exercise outside the competition and publishes the results.
Claimed by
DARPA and competing teams, November 2026
Open
01
A public lost-person scoreboard
A free, public test of the maps that predict where a lost person will be found. It needs a set of closed cases, each with where the person was last seen, their category (hiker, child, person with dementia and so on) and where they were found, with names removed, plus code that scores every method the same way: Koester's distance rings; agent-based simulations, which send thousands of simulated walkers across a terrain model; or an AI model asked to rank search areas. It uses closed cases only and touches no live search.
The briefCredit · unclaimed02
Sweep widths for drone searches
Search planners turn the ground a searcher has covered into a chance of having found the person with a number called the effective sweep width, a single measure of how well a searcher, aircraft or drone detects a given object in given conditions. Measure it for camera drones on a test course: life-size mannequins, warmed for thermal cameras, placed at surveyed spots in grass, scrub and woodland; drones flown at set heights and speeds; and detections by a person watching the footage and by detection software, each scored against the true positions. Fly only over training ground with the owner's permission, with a pilot who holds whatever certificate local drone rules require, never in a live search.
The briefCredit · unclaimed03
Read rescue marks from photos
An app that reads, from a phone photo, the marks rescue teams paint in orange on searched buildings under the INSARAG system: victim marks, a large V with counts of live and dead people such as "L-2" or "D-3" that are crossed out and rewritten as people are removed, and clearance marks, a diamond with C for clear or D for deceased only, plus the team code and date. It turns each mark into the fields of the team's coordination record and flags marks that disagree with the record. Train and test it only on marks painted on boards and walls at training sites.
The briefCredit · unclaimed
Open 01 · the brief
A public lost-person scoreboard
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free, public test of the maps that predict where a lost person will be found. It needs a set of closed cases, each with where the person was last seen, their category (hiker, child, person with dementia and so on) and where they were found, with names removed, plus code that scores every method the same way: Koester's distance rings; agent-based simulations, which send thousands of simulated walkers across a terrain model; or an AI model asked to rank search areas. It uses closed cases only and touches no live search.
Why it is still open
The largest case collection, ISRID, held 532,816 incidents by Koester's October 2025 count and is available on request from him. MapScore, the public scoring portal built at George Mason and Brigham Young universities, was "suspended due to not much use", and its public files hold only the Arizona, New York and Yosemite cases out of about 400. The University of Glasgow said in April 2025 that its simulated-walker model needs "further development work and validation" before real use, and Undark reported in 2024 that there are few independent studies of how well lost-person methods work. SAREnv, an open benchmark published in September 2025 by Kasper Grøntved and colleagues, scores drone search paths against 60 maps built from lost-person statistics, not against where real people were found.
github.comgla.ac.ukapi.crossref.orgundark.orgalpine-rescue.org
Built with what exists
MapScore's open code and public cases; the simulated-walker method published by Jan-Hendrik Ewers and colleagues at Glasgow (arXiv 2412.13317); free terrain, trail and stream data such as OpenStreetMap; a language model to pull fixed fields out of written incident reports, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); Python.
First step
Download MapScore's public cases, rerun its scoring on simple distance rings, and ask two rescue organizations whether they will share closed-case locations with names removed.
Done looks like
At least 300 closed cases from at least three countries, none of them used to build the methods being scored, with every method scored on the same measure and the cases, scores and code posted so that anyone can add a method and get a score.
Open 02 · the brief
Sweep widths for drone searches
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Search planners turn the ground a searcher has covered into a chance of having found the person with a number called the effective sweep width, a single measure of how well a searcher, aircraft or drone detects a given object in given conditions. Measure it for camera drones on a test course: life-size mannequins, warmed for thermal cameras, placed at surveyed spots in grass, scrub and woodland; drones flown at set heights and speeds; and detections by a person watching the footage and by detection software, each scored against the true positions. Fly only over training ground with the owner's permission, with a pilot who holds whatever certificate local drone rules require, never in a live search.
Why it is still open
A 2023 paper by Alexander Sulaiman and Ken Pierce notes that sweep-width tables for aircraft and ships appear in manuals such as the IAMSAR Manual, the international search and rescue manual, but "no such tables exist for drones", and their model of helicopter sweep widths did not yet match the published tables. Open sweep-width experiments exist for people searching on foot, such as forensic searches at the University of Toronto Mississauga from 2023 to 2025, and drone trials report the share of targets found, such as a 2025 test with 78 volunteers in Mediterranean karst and Norway's SAVIOUR 2024 exercise, where six helicopter and six drone crews flew 48 sorties to find 251 people (a preprint, April 2026), but searches of the web, arXiv and GitHub found no published sweep widths for drones.
Built with what exists
a commercial drone carrying thermal and visible cameras; open detection models such as the YOLO family; the open Sweep Width Calibration Module, written in R, which turns detection records into sweep widths; survey-grade GPS to place the targets.
First step
Lay out 30 mannequin spots on open grass at a drone training field, fly one height and speed, and compute a first sweep width with the open module.
Done looks like
A table of sweep widths with confidence ranges for at least three heights, three kinds of vegetation, and day and night, each cell from at least 100 target placements, for both human reviewers and software, published with the raw detection records.
Open 03 · the brief
Read rescue marks from photos
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An app that reads, from a phone photo, the marks rescue teams paint in orange on searched buildings under the INSARAG system: victim marks, a large V with counts of live and dead people such as "L-2" or "D-3" that are crossed out and rewritten as people are removed, and clearance marks, a diamond with C for clear or D for deceased only, plus the team code and date. It turns each mark into the fields of the team's coordination record and flags marks that disagree with the record. Train and test it only on marks painted on boards and walls at training sites.
Why it is still open
INSARAG's review of the response to the earthquake in Türkiye on 6 February 2023, in which about 5,000 people from 90 international teams took part, found that building marks were sometimes not used, were hard to apply on badly collapsed buildings and were changed by local people. It found that the information in ICMS, INSARAG's coordination software, "does not exactly match the activities that occurred", and it recommended more use of photos. Searches of the web and GitHub found no tool that reads these marks from photos.
Built with what exists
phone cameras; an AI model that reads images, such as Gemini or Claude, or an open detector from the YOLO family with text recognition; Annex B26 of the INSARAG Guidelines, which defines the marks, as the specification.
First step
Paint 50 marks on plywood boards following Annex B26, photograph each from three angles, and test one image model on them.
Done looks like
1,000 photos of marks painted by trainees at three or more training sites, in different light, angles and states of wear, each transcribed by two people; the app reads at least 95% of fields correctly, and every misread count of live or dead people is listed in the published results.