Airbus ATTOL, the Vision-Based Automatic Takeoff That Ignored the ILS and Read the Runway With a Camera

In December 2019, an Airbus wide-body airliner took off from Toulouse-Blagnac, France, with both pilots’ hands off the controls - and it did it without an instrument landing system (ILS). Instead of following radio beams, the aircraft read the runway with cameras and computer vision, tracking the centerline through the wind and rotating itself into the climb. This was ATTOL (Autonomous Taxi, Takeoff and Landing), an Airbus UpNext demonstrator that proved a machine could navigate a runway the way a human pilot does: by seeing it.

What Was Airbus ATTOL?

ATTOL stands for Autonomous Taxi, Takeoff and Landing. It was a technology demonstrator run by Airbus UpNext, the company’s wholly owned innovation subsidiary, to test whether an aircraft could handle ground and low-altitude operations using image recognition rather than ground-based radio guidance.

The core idea was simple to state and hard to build. Cameras mounted on the aircraft fed a computer vision system trained to recognize the runway, its markings, its edges, and its centerline - the same visual cues a human pilot relies on. The airplane processed those images in real time and steered accordingly.

The first milestone came on December 18, 2019. Over two flights spanning about four and a half hours, the aircraft performed eight fully automatic takeoffs. Two pilots sat up front as safety backups, but nobody touched the sidestick until the airplane was well into its climb.

How Is Vision-Based Automation Different From Autoland?

Automatic landing is not new. Category III autoland - the triple-redundant autopilot that brings an airliner down in fog too thick to see through - dates back to the 1960s. British European Airways was performing automatic landings in the Hawker Siddeley Trident decades ago.

But autoland has never actually seen the runway. It follows a pair of radio beams: a localizer for left-right alignment and a glideslope for the approach angle. The autopilot simply keeps itself centered in those beams down to the flare.

That system is proven and has saved countless flights, but it has hard limits:

It only works where an ILS is installed, tuned, calibrated, and protected.
It covers the approach and landing only - there is no beam for takeoff.
The ground equipment is expensive, requires protected zones, and can be disturbed by something as ordinary as a truck parked in the wrong place.

ATTOL asked a different question: what if the airplane didn’t need the ground equipment at all? What if it could just look at the runway markings that every paved runway already has?

Why Perception Is the Hard Part

Making an airplane hold a heading and pitch attitude automatically is old news - we’ve done it for roughly a century. The genuine frontier problem is perception.

The data shows how much work that took. Airbus flew more than 500 test flights for the program, but most of them weren’t autonomous maneuvers. The majority were spent gathering raw video - hours of runway footage in different light, weather, and angles - so engineers could train the system to recognize a runway from every realistic perspective.

The difficulty is teaching a computer to look at a smear of gray asphalt in low winter sun, streaked with rubber skid marks and cut by a taxiway intersection, and reliably conclude: that is the runway, that is the centerline, and I am three feet left of it. Humans do this effortlessly. Machines find it extraordinarily hard - the same perception challenge the self-driving car industry has wrestled with for years. Perception is always harder than control.

The Three Phases: Taxi, Takeoff, and Landing

ATTOL was built around three capabilities, all driven by what the aircraft could see rather than what a ground station broadcast:

Takeoff - demonstrated first, in December 2019.
Automatic taxi - the aircraft guiding itself along taxiways using its cameras, added in the first half of 2020.
Vision-based landing - the aircraft reading the runway on final and flaring itself, again with no ILS beam involved.

Airbus wrapped up the full program in mid-2020.

Why Airbus Refused to Oversell It

The most notable part of this story is what Airbus didn’t claim. Demonstrating a wide-body taking off with the pilots’ hands in their laps is exactly the kind of moment a company might use to announce that pilotless airliners are imminent. Airbus did the opposite.

When the program closed, its own project leads stated plainly that the goal was never to move toward autonomous airliners flying passengers with nobody up front. The stated aim was to explore whether autonomous technologies could assist pilots - reducing workload, improving safety, and letting crews focus on decisions that matter instead of the mechanical grind of tracking a centerline in a gusty crosswind.

The Real Pros and Cons

On the promise side, vision-based automation is powerful precisely because it needs no infrastructure. Only a small fraction of the world’s runways have a full ILS on both ends. An aircraft that can land using nothing but a camera and standard runway markings could extend automatic-landing capability to thousands of airports that will never receive an ILS - a safety argument, not just a convenience. A crew facing low fuel and marginal weather at a small field would gain a tool they simply don’t have today.

The takeoff capability matters more than it sounds. Takeoff and initial climb are among the highest-workload, least-forgiving phases of any flight, and an engine failure at rotation is one of the hardest things a crew ever handles. Automation that holds the aircraft precisely on centerline through that moment could free the pilots’ attention for the go/stop decision - backing up the crew, not replacing it.

On the other side of the ledger, vision systems fail differently, and arguably worse. A disturbed ILS beam usually degrades in a way the system can detect and flag - it knows it’s confused. A computer vision system can be confidently wrong. Faced with a snow-covered runway, blinding glare, or confusing construction paint, it can produce an answer that is clean, precise, and completely incorrect. That failure mode - high confidence, total error - is what keeps safety engineers up at night, because the system never raises its hand to say it’s lost. It just steers you wrong while looking certain.

Then there’s certification. ATTOL was a demonstrator, flown in a controlled environment with test pilots ready to take over instantly. Certifying automation like this for paying passengers means proving it’s safe not on a good day in Toulouse, but on the worst day at the worst airport in the worst light - bounding every failure mode and proving the perception system can’t be fooled. That is years of work, and for full autonomy it arguably isn’t solvable with today’s technology. That’s precisely why Airbus framed ATTOL as pilot assistance, not pilot replacement.

What Happened Next: DragonFly

ATTOL didn’t end as a dead end - it fed directly into Airbus’s follow-on program, DragonFly, which points the vision technology at concrete pilot-assistance jobs:

Automated emergency diversion - if the crew is incapacitated, the aircraft can select a suitable airport, navigate to it, and communicate with air traffic control and the airline.
Automatic landing using the same camera-based approach ATTOL pioneered.
Assisted taxi - guidance and audio alerts as crews move around busy, unfamiliar airports at night, where many ground collisions happen.

The philosophy is clear: not “build a robot pilot,” but “take the hardest, highest-risk moments and give the human crew a capable backstop.” It shares a lineage with Garmin’s emergency autoland systems in light aircraft, where a passenger can push one button to land the airplane if the pilot is incapacitated. Different scale, same idea - automation as a safety net that deploys when everything has gone wrong, not a replacement that flies when everything is fine.

Who Else Is Working on This?

Airbus, through Airbus UpNext, is the name on ATTOL and DragonFly, but the whole industry is racing on the same core problem: machine perception you can trust with a life. Boeing has its own autonomy efforts. Defense programs have pursued vision-based landing for years for carrier operations and austere fields. And every eVTOL air taxi company is betting long-term on some version of this technology, because the economics won’t support putting a trained airline pilot in every four-seat flying taxi.

The Bigger Shift

For sixty years, the answer to landing in bad conditions was to make the ground smarter - put a beam on the runway and teach the airplane to follow it. ATTOL flipped that. It made the airplane smart enough to not need the beam, to simply look at the runway and understand it.

That’s a profound and probably correct long-term direction. But the program earns respect because its builders told the truth about how far it still has to go: they flew a wide-body off the runway with nobody on the controls, then stated openly that this is meant to help pilots, not replace them, and that full autonomy is not around the corner. The airplane that reads the runway with a camera is real, it flew, and it works on a good day. The future of cockpit automation is the long, unglamorous grind of making it work on the bad ones.

Key Takeaways

Airbus ATTOL (Autonomous Taxi, Takeoff and Landing) demonstrated camera-based automation, performing 8 automatic takeoffs on December 18, 2019 with no ILS beam.
Unlike Cat III autoland (which follows radio beams), ATTOL used computer vision to read runway markings directly - the same cues human pilots use.
Airbus flew over 500 test flights, most spent gathering runway video to train the perception system; the program covered taxi, takeoff, and landing before wrapping up in mid-2020.
The key risk is that vision systems can be confidently wrong, a failure mode harder to detect than a degraded radio signal - a major certification hurdle.
ATTOL fed into DragonFly, which applies the technology to pilot-assistance tasks like emergency diversion and assisted taxi, framing autonomy as a safety backstop, not pilot replacement.