[gregmac]

> Develop software that can fly the planes autonomously from strip to strip (I assume this is the really hard part, but I am under the impression that autonomous flying is a much easier problem than autonomous driving?).

As a software developer, this part is what I don't like.

We're still quite a ways from fully autonomous driving cars (as in: don't rely on a human taking over for backup). A bad bug in an autonomous car could drive you at high speed into a wall, but there can at least be an "emergency stop" button that disables the main processor and jams on the brakes.

Planes have no such ability to just "stop". At best, they could deploy a parachute, but even then landing safely is by no means guaranteed.

I think we need a decade or so of fully autonomous cars being accepted into daily life before this can be attempted with anything that flies.

[holycrapwhodat]

As both a software developer and a pilot...

We're (much?) closer to Fully-Self-Flying planes than FSD cars because the problem space is - perhaps counterintuitively - MUCH smaller to tackle. And we have a lot more experience tackling it.

Additionally there could easily be remote pilots as backup in case of catastrophe (See remote piloted military and border patrol UAVs)

And pulling a parachute at 1000'+ altitude actually has quite a bit of precedent (See https://en.wikipedia.org/wiki/Cirrus_Airframe_Parachute_Syst...)

Now... There's probably a lot of cultural and regulatory reasons why the "string of automated glider ports" idea will never come to fruition.

But... As far as technical hurdles go, there's not much new technology that would need to be invented here.

[theYipster]

As an industry leader on this specific subject matter... I agree: the problem space is much smaller, in theory.

However, certification requirements and safety assurance needs will drive both cost and time into realizing fully autonomous aircraft. They will be here, but we are 10-15 years away.

The problem is in how to certify machine learning code. Today, you can't. Existing AMCs (accepted means of compliance) are incompatible with the nature of ML. (The breakdown is specifically with assurance architectures focused on code traceability and coverage.) A new architecture for demonstrating safety assurance with AI/ML is needed, and is being built, but is still 1.5-2 years away from being released, and then it will take another year or two before a CAA (civil aviation authority, like the FAA or EASA) will certify a component with ML code--and that will not be an autonomous pilot. That will come in time, but the industry is conservative--especially on safety-critical matters--and it will take years to develop trust in both technology, human factors, and methodology to work up to autonomously flown passenger aircraft.

From the regulator perspective, EASA has taken poll position in thought leadership. Google their AI Roadmap or their Concept Paper for Level 1 Machine Learning Applications.

[silvestrov]

> string of automated glider ports

There is a small precedent here: the space shuttle flew almost 100% by computers.

There was a small involvement of the pilots when landing and some change of software due to small RAM size in the computers.

I could imagine "Fully-Self-Flying planes" would start out with cargo planes between areas with low population.

[topspin]

Point of information; certified autoland systems appeared on airliners in 1968. Yes, we're a ways off from the level of automation for a fully pilotless system as proposed here, but I think it a matter of a few years, perhaps a decade.

[jonfw]

You know what they say- the last 20% of the problem is 80% of the effort. I think that these percentages probably understate things, but the point is that getting something to be mostly functional in ideal circumstances really isn't much of an achievement.

[topspin]

"mostly functional in ideal circumstances"

Autoland was developed for and is usually used in poor conditions.