Aircraft

AI autopilot could let autonomous aircraft fly in busy airspaces

In a test of the artificially intelligent autopilot, both it (left) and a human pilot (right) controlled virtual aircraft in the same simulated airspace
Carnegie Mellon University
In a test of the artificially intelligent autopilot, both it (left) and a human pilot (right) controlled virtual aircraft in the same simulated airspace
Carnegie Mellon University

While pilots commonly use an autopilot when cruising at high altitudes, they typically switch to manual control when entering crowded lower airspaces. However, what if the plane has no pilot? Well, a new AI autopilot system for busy airspaces may be the answer.

Currently being developed by researchers at Carnegie Mellon University, the system was trained on data collected at Pennsylvania's Allegheny County Airport and Pittsburgh-Butler Regional Airport.

Along with utilizing the autonomous aircraft's existing instrumentation – and information provided by local air traffic controllers – it incorporates six cameras and a computer vision system. According to the scientists, the latter setup allows it to visually spot nearby aircraft much as a human pilot would do.

The system is able to subsequently track those aircraft and predict their trajectories, taking action to avoid collisions when necessary. It can reportedly even understand plain-language spoken radio communications from other pilots or from airports, responding with a synthetic speech system of its own.

Although the technology has yet to be tested on an actual airplane, it has been trialled in a setup that involved two linked flight simulators. In that setup, a human piloted one virtual aircraft, while the AI system piloted another in the same airspace. It was found that the AI was able to successfully avoid collisions with the human-piloted aircraft, even when the pilot had little experience at flying a plane.

The scientists hope that once developed further, the system could be applied to autonomous aircraft such as air taxis or delivery drones.

"This is the first AI pilot that works in the current airspace," said team member Assoc. Prof. Sebastian Scherer. "I don't see that airspace changing for UAVs [unmanned aerial vehicles]. The UAVs will have to change for the airspace."

Source: Carnegie Mellon University

  • Facebook
  • Twitter
  • Flipboard
  • LinkedIn
6 comments
Daishi
I don't have any interest in getting in a plane without a human pilot or even flying near one that doesn't. I don't know the experiences people have with technology that convince them this is a good idea but I find it fickle and buggy. The more complexity you bring in to solve corner cases the more difficult to predict or replicate interactions you create through complexity. Automating teaching college courses seems like lower hanging fruit maybe they should try that instead. They can restart or power cycle the professor when it stops responding or provides incomprehensible data instead of having a few hundred people crashing to their death.
Eggbones
Tactical deconfliction is a bad idea that would be replaced by 4-dimensional trajectory management as soon as it's available. The sky needs to be comprehensively mapped and managed - measures to stop things bumping into each other are short-sighted.
sidmehta
Hope this works better than Tesla, which couldn't see a tractor-trailer standing _sideways_ in broad daylight. And rammed straight in to it,
RaoulToday
I don't know why you bother reading these articles. Cutting edge always means a little blood. I would like to sit around a table with you all and Chuck Yeager, while you guys told him how crazy it was to break the sound barrier, and go faster and farther than ever before.
TpPa
I don't want to hear about any complaints about drones anymore, I'd much rather have a Mavic 3 land on my head than a plane that the AI screwed up in
TpPa
A nice sunny day reflecting off a shiny colored semi trailer makes a lot of the sensors useless, they need contrast to comprehend something is there, LiDAR ? unsure about.