BRENNEN

Industries / Transport and vehicles / Aviation

Air traffic control

Air traffic control is the work of keeping aircraft safely apart and moving them in an orderly flow, mostly through radio instructions from controllers on the ground.

6

Milestones achieved, from 1920 to 2019.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1920

    The first control tower

    On 25 February 1920 the UK Air Ministry commissioned an "Aerodrome Control Tower" at Croydon, then London's airport, raised 15 feet above the ground with large windows on all four walls. CANSO, the trade body of air navigation service providers, calls it the world's first air traffic control tower. Its staff, called Civil Aviation Traffic Officers, gave pilots traffic, position and weather information by radio and tracked about a dozen daily flights with pins and flags on paper maps.

    canso.org

    Credit

    UK Air Ministry

  2. 1956

    Collision over the Grand Canyon

    On 30 June 1956 a TWA Super Constellation and a United Airlines DC-7, which had left Los Angeles about three minutes apart, collided over the Grand Canyon and all 128 people aboard died. Both had left the airways for direct routes under visual rules; large parts of the US had no radar coverage, and controllers relied on pilots' position reports. Congress then funded modernization: the Civil Aeronautics Administration ordered 23 long-range radars for $9 million, and the Federal Aviation Act of 1958 created the Federal Aviation Agency.

    Lessonwith no radar watching, keeping aircraft apart was left to pilots' eyes, and that failed.

    aopa.org

    Credit

    Civil Aeronautics Board (probable cause: "pilots did not see each other in time"), US Congress

  3. 1981

    Striking controllers fired

    On 3 August 1981 members of the Professional Air Traffic Controllers Organization (PATCO), a US union, went on strike, which was illegal for federal workers; sources put the strikers at about 13,000 (ABA Journal) or over 12,000 (UTA Libraries). On 5 August President Ronald Reagan fired the 11,345 who had not returned (ABA Journal's figure). Supervisors, non-strikers and military controllers restored traffic to about 75% of normal within two days (ABA Journal); UTA Libraries counts about 3,000 supervisors, 2,000 non-striking controllers and 900 military controllers, with 7,000 flights cancelled.

    Lessona controller workforce takes years to rebuild; UTA Libraries says the shortage lasted into the George H. W. Bush administration (1989 to 1993).

    abajournal.comlibraries.uta.edu

    Credit

    Reagan administration and the FAA

  4. 2011

    One network manager for Europe

    After the April 2010 eruption of Iceland's Eyjafjallajökull volcano, EU Regulation 677/2011 (7 July 2011) said the events had "demonstrated the need to create a central entity" to coordinate responses to aviation crises. It set up a single Network Manager for European air traffic, in charge of route network design, radio frequencies, transponder codes and air traffic flow management (spacing out departures so no stretch of airspace gets more traffic than its controllers can handle), plus a European Aviation Crisis Coordination Cell. EUROCONTROL, already entrusted with flow management by its member states, runs the Network Manager; its own page says its seasonal planning draws on data from about 350 airlines, 68 area control centers and 43 states.

    Lessona crisis that crosses many countries' airspace needs one body that can coordinate all of it at once.

    eur-lex.europa.eueurocontrol.int

    Credit

    European Commission and EUROCONTROL

  5. 2015

    A tower run from afar

    On 21 April 2015 LFV, Sweden's state-owned air navigation service provider, began controlling takeoffs and landings at Örnsköldsvik Airport from a remote tower center in Sundsvall. Saab's sensors at the airport send live data to the controllers there; Saab puts the distance at about 100 km, AIN at about 78 miles (about 125 km). The Swedish Transport Agency had given final operational approval by 3 November 2014, when Saab said Örnsköldsvik would become "the world's first airport remotely controlled by Remote Tower Services."

    ainonline.comsaab.com

    Credit

    LFV and Saab

  6. 2019

    Tracking aircraft from space

    Over the ocean, beyond ground radar, aircraft on the North Atlantic were kept 5 to 10 minutes apart along their track (NAV CANADA's figures). From March 2019 Aireon's ADS-B receivers, carried on 75 Iridium NEXT satellites (66 working, nine spares), gave controllers live positions there; ADS-B is a signal in which each aircraft broadcasts its own satellite-derived position. NATS, NAV CANADA and NAV Portugal then trialled cutting the spacing between aircraft following each other from 40 nautical miles to 14, and after 19 months the standard was adopted permanently in November 2020 and published in ICAO documentation.

    navcanada.caadsgroup.org.uk

    Credit

    Aireon, NAV CANADA, NATS and NAV Portugal

Watch for

  1. A rebuilt US control system

    In July 2025 Congress gave $12.5 billion as a "down payment" on the FAA's Brand New Air Traffic Control System, and on 4 December 2025 the FAA picked Peraton as prime integrator, the company that coordinates design, testing and rollout. The US Department of Transportation (DOT) and FAA say they will finish the overhaul by the end of 2028 and that about $20 billion more is needed. At an April 2026 briefing they reported nearly half of old copper lines converted to fiber, about 270 radio sites converted, more than 40 new voice switches installed, and contracts with RTX and Indra to replace up to 612 radars by June 2028.

    Achieved whenan outside reviewer, such as the DOT Inspector General or the Government Accountability Office, confirms the new radars, telecommunications and automation are in service.

    hklaw.com

    Claimed by

    US DOT and FAA, December 2025 and April 2026

  2. Software that spaces out flights

    The FAA has launched SMART (Strategic Management of Airspace, Routes and Trajectories), an $875 million software system built by Air Space Intelligence, at three airports serving Washington, D.C. It weighs planned airline flights, weather and wind, and adjusts schedules to reduce congestion before it builds. Transportation Secretary Sean Duffy says it will cut delays and controller stress; the controllers' union, NATCA, says it was not involved in its design, testing or rollout.

    Achieved whendelay data reviewed by someone outside the FAA show fewer delays at the SMART airports than before, with no rise in safety events.

    wamc.org

    Claimed by

    US DOT and FAA, September 2026 (reported by NPR)

Open

  1. 01

    Public ATC transcription scoreboard

    A fixed test and public scoreboard for turning recorded ATC radio into text, using only audio anyone can download, with one scoring script for every system and new speech-to-text models added as they appear.

    The brief
    Credit · unclaimed
  2. 02

    Real-traffic conflict test set

    A shared set of conflicts (two aircraft predicted to come closer than the minimum allowed spacing), cut from recorded traffic and replayed in the BlueSky open-source ATC simulator, with one scoring script, so that rule-based, reinforcement-learning (software that learns by trial and error in simulation) and language-model solvers can be ranked on the same cases. Simulation only.

    The brief
    Credit · unclaimed

Open 01 · the brief

Public ATC transcription scoreboard

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A fixed test and public scoreboard for turning recorded ATC radio into text, using only audio anyone can download, with one scoring script for every system and new speech-to-text models added as they appear.

Why it is still open

Papers report word error rates (the share of words a system gets wrong) on different splits and training data, so they cannot be compared directly. A January 2026 paper from Université de Lorraine notes that "only one hour of transcribed audio is freely available" in the ATCO2 corpus; on it, an off-the-shelf Whisper-small model scored 63.32% word error rate and a fine-tuned one 19.54%, while the baselines it cites used other setups. A 2020 benchmark by Zuluaga-Gomez and colleagues reported a 7.75% average across four databases but published no leaderboard. A search for a maintained ATC leaderboard found none.

arxiv.orgarxiv.org

Built with what exists

the free one-hour ATCO2 test set; the ATCOSIM corpus (about 10 hours of simulated controller speech from Graz University of Technology and the EUROCONTROL Experimental Centre, posted on Hugging Face); OpenAI's open Whisper models; ElevenLabs Scribe v3 (see the Note "Fewer word errors"); a Hugging Face Space to host the table.

  1. First step

    Download the ATCO2 one-hour test set and ATCOSIM, write normalization rules for numbers, callsigns and the spelling alphabet, and score Whisper on both.

  2. Done looks like

    At least five systems scored on the same audio with one published text-normalization script, and one outside person reproducing a listed score within 0.5 percentage points.

Open 02 · the brief

Real-traffic conflict test set

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A shared set of conflicts (two aircraft predicted to come closer than the minimum allowed spacing), cut from recorded traffic and replayed in the BlueSky open-source ATC simulator, with one scoring script, so that rule-based, reinforcement-learning (software that learns by trial and error in simulation) and language-model solvers can be ranked on the same cases. Simulation only.

Why it is still open

The authors of BlueSky-Gym (TU Delft, 2024) wrote that "to date there is no common basis on which these RL techniques are compared"; their seven environments cover vertical descent, obstacle avoidance and traffic merging, not conflicts drawn from real traffic. A search for a shared real-traffic benchmark scoring reinforcement-learning and language-model solvers side by side found none.

repository.tudelft.nl

Built with what exists

OpenSky Network ADS-B history; the traffic Python library to clean tracks; BlueSky (MIT license; it includes conflict detection and resolution and plays timestamped scenario files) and BlueSky-Gym; any open or hosted language model for a language-model baseline.

github.com

  1. First step

    Take one day of OpenSky data for one busy upper-airspace sector, find aircraft pairs predicted to come within 10 nautical miles of each other at the same level, and write 20 of them as BlueSky scenario files.

  2. Done looks like

    At least 500 scenarios from at least three regions, published with the scoring script; scores (losses of separation, extra distance flown, number of instructions) for at least three solvers, including BlueSky's own conflict resolution; a second team reruns one solver and gets the same scores.