Crash at Reagan National Airport, DC. Small aircraft down in the Potomac.

kyleb said:
Engineering i.e. hardware design is a vastly different undertaking than designing a system that has to operate with a ton of other systems that are impacted by things not necessarily in their direct control.

Again, the example of self driving cars comes to mind.
Click to expand...
Ok, I don't think you're contradicting the quoted post. Certainly you recognize that SpaceX hardware operates with a ton of systems that are impacted by things but in their direct control. I don't think anyone would claim that everything between firing a rocket into space and catching it back at the launch pad or landing it on a floating barge, not to mention docking and recovering a spacecraft with the space station, is within SpaceX's control.
 
Self driving vehicles are a much better analogy to ATC AI than anything SpaceX does. It's not a particularly difficult software problem, and could be solved using current AI tools. It just takes a lot of development resources and time to get performance to a level acceptable for such a critical system, with 90% of that effort spent on edge cases. Right now, there is little impetus in either the government or private sector to expend those resources on that problem. As AI tools continue to improve, eventually the problem will become solvable at a reasonable cost, and it will happen. Someday.
 
You don't have to "blow things up" to train an AI for ATC.

Tesla's FSD computer, the one in each car, is calculating a driving solution all the time, even when Autopilot is not engaged. When the human driver's actions diverge significantly from FSD's solution, the data is sent back to Tesla and, when the human's performance was better than FSD, is used for AI training.

That is one way that an ATC AI would be trained.

One way in which FSD has an advantage over human drivers is that it is watching all eight exterior cameras simultaneously without ever being distracted or having to "turn its head" in different directions to see all necessary angles. The ATC AI would have a similar advantage in that it would have more information available to it than any single controller now has.

For example, it would know about all of the other airplanes currently flying toward a particular STAR and at what time they will reach each waypoint. Airplanes hundreds of miles apart can be sequenced and deconflicted hours before they arrive at the conflict point creating a more efficient system, with fewer bottlenecks, than is currently possible.

FSD has a features with some parallels to this sequencing in its navigation software. FSD knows when you will need to charge and what Superchargers are available along your route in the area where you will be ready to charge. It also knows how many stalls are available at each charging site, how many other cars are navigating to those chargers, and at what times they will all arrive. It uses this data to spread the charging demand across the stations to avoid anyone having to wait for a charger.
 
Bob Noel said:
I don't think the collision avoidance algorithm is quite as trivial as you might think. It's not difficult to come up with realistic scenarios where escape maneuvers create even more collision hazards.

(then there is the whole problem of dealing with spoofed ADS-B messages).
Click to expand...
It is every bit as "trivial" (and by that, I don't mean easy for any coder to write, far from it) as similar stuff exists today.* There are far fewer items to predict (despite how busy the ADS-B map looks) than what is tracked in various industries. And we aren't working with entirely random dots on the screen. In fact, the only real issue is that the pilot must obey the alerts, which we already know ain't a 100% certainty.
*Former DoD programmer, who worked on image recognition "back in the day", and AI in the private sector, so I'm likely looking at it all wrong. It could have been done decades ago. And should have. ADS-B is what I would call a "kluge" when it comes to avoidance.
 
I'll note that we aren't at a zero rate for mid-air collisions among aircraft under human ATC control today, so there's no reason to hold AI ATC to that standard.
 
Lindberg said:
I'll note that we aren't at a zero rate for mid-air collisions among aircraft under human ATC control today, so there's no reason to hold AI ATC to that standard.
Click to expand...
When was the last mid-air that was caused by an ATC error or failure? I can't think of anything recent.
 
You must log in or register to reply here.
Back
Top