Will computers ever fly alone?

Aviation analyst Brian Foley stirred the pot in a recent article published by Forbes asserting that Boeing’s 20-year outlook for pilot and technician demand overestimates future demand for pilots by discounting the effects of automation advances. Boeing predicts 763,000 pilots will be needed worldwide over the coming 20-year period, a 5-percent reduction from the firm’s 2019 forecast. The Seattle manufacturer expects airline layoffs and furloughs prompted by the coronavirus pandemic will eventually be offset by resurgent demand, crew retirements, and expanding aircraft fleets. Canadian flight simulator maker CAE issued its own forecast November 9 predicting demand for 260,000 new pilots through 2029, including 27,000 new pilots needed by the end of 2021.

CAE also noted that many of the recently furloughed pilots have found work in other industries as airlines scale back routes amid plunging demand for air travel (which seems unlikely to recover as long as the coronavirus remains a factor). “On the other hand, data indicates that the industry will face significant challenges in the upcoming years to meet the demand for pilots,” wrote Nick Leontidis, CAE group president of civil aviation training solutions, in the online presentation of the firm’s forecast. 

“That may be, but only if the world of cockpit technology stands still for the next two decades, which would seem highly unlikely,” Foley wrote, specifically in reference to Boeing’s prediction, though Boeing and CAE expect similar future demand for pilots.

Foley noted that automation has already arrived, to a degree. Garmin’s Autoland system, unveiled in 2019, can fly a GA aircraft to the nearest suitable runway and land if the human pilot becomes incapacitated—a passenger just has to press one button. Military aircraft have flown without a human aboard for several years, with drones high overhead and helicopters hugging the earth to deliver cargo drops in Afghanistan and elsewhere. Autonomous tech startup Xwing has secured FAA Part 135 air carrier certification for unmanned deliveries in a converted Cessna Grand Caravan, and hopes to begin autonomous cargo delivery operations in 2022, supervised by humans on the ground, to start.

Urban air mobility could drive further progress in automation, or, at least, demand for progress. Some firms vying for a slice of the future air taxi service pie are aiming for autonomy sooner than later, though others envision human pilots will be needed, at least for a few years, to achieve regulatory approval and public confidence.

Foley did a little math based on Boeing’s predicted demand for 48,400 new airliners by 2039, assuming that it takes about 10 pilots (five two-pilot crews) to staff each aircraft.

“An airline’s chief focus is on profitability, and finding ways to cut costs by downsizing the number of pilots needed to operate an aircraft is a strong motivator,” Foley wrote. “This was proven long ago when technology first made the navigator and later the flight engineer obsolete, saving airlines a considerable amount of money.”

If automation can enable single-pilot operation of 10 percent of the 48,400 airliners that are expected to be sold, that would reduce the bottom line demand for pilots by 10 percent, nearly 50,000 pilots, Foley wrote: “Factoring in additional airline pilots scuttled by automation prior to 2039, plus the technological culling of the copilot ranks in the civil business jet and helicopter fleets, suggests Boeing’s forecast is overstated by perhaps 10 [percent] and maybe more.”

Pilots quickly pushed back after Foley posted his prediction on LinkedIn, with more than one commenter opining that it will “never happen.”

In a follow-up email exchange, Foley noted that his “thesis was centered on replacing only a portion of copilots by 2039. Full autonomy could still possibly occur in the cargo and military arenas during that timeframe.”

‘Banger’ bested

Before we start punching holes in Foley’s argument, here’s one more indication of computer capability that he did not mention: In August, around the time Xwing made a few headlines with its highly automated Caravan cargo flights, the Defense Advanced Research Projects Agency (DARPA) hosted the AlphaDogfight Trials, pitting a highly trained human against artificial intelligence in a series of simulated dogfights. According to Air Force Magazine, the unnamed human was a District of Columbia Air National Guard pilot with more than 2,000 hours in the General Dynamics F–16, and a graduate of the U.S. Air Force Weapons School F–16 Weapons Instructor course, call sign “Banger.”

An AI algorithm developed by Heron Systems made short work of Banger, notching five victories in their five head-to-head matchups. The computer was programmed to work within human limits, so it did not “pull” excessive Gs, though it ignored other limitations that the Air Force imposes on real-world fighter pilot practice, such as maintaining at least 500 feet of separation.

Banger’s human brain needed more time than the computer to make calculations, decisions, and adjustments. While Banger learned and improved in successive matchups, the human pilot did not manage to put so much as a single simulated round on target (this was a guns-only dogfight).

Heron Systems earned its showdown with humanity by beating other AI entrants, one fielded by Aurora Flight Sciences, a longtime leader in aviation automation that became a Boeing subsidiary in 2017, and others, on its way to beat up Banger. The one-sided nature of the exercise came as no surprise to Kenneth Goodrich, deputy manager of NASA’s Advanced Air Mobility project. 

Goodrich, whose NASA career began more than 35 years ago and included work for the Advanced General Aviation Transport Experiments consortium that led to several safety-enhancing advancements, recalled working on dogfight logic decades ago. Computers were pretty good at winning tactical engagements even then, and Goodrich said he is not at all surprised that Heron's AI program (call sign “Falco” for the DARPA competition) cleaned the human pilot’s clock in the confines of a one-on-one tactical engagement.

“It is probably the sort of task that computers are very capable of outperforming humans,” Goodrich said. But there is a big distinction between winning the dogfight and winning the battle, or the war, Goodrich said.

“From the perspective of the fighter pilot, the big challenge is how you initially set up and get into engagements,” Goodrich said. “Once you’re actually in the furball, definitely the computer logic, the speed, the accuracy, has a lot of advantages.”

But those advantages are primarily tactical, not strategic. And in the world of civilian air transportation, there’s a huge difference between the most sophisticated autopilot of the present day, and a system able to assume responsibility for every phase of flight, including emergencies and other situations that introduce “unknown unknowns” to the aeronautical equation.

Think about the decision that Capt. Chesley "Sully" Sullenberger made to land US Airways Flight 1549 in the Hudson River, rather than attempt to glide to the nearest runway in a densely crowded metropolis, or the heroism of the crew of United Airlines Flight 232, who saved 184 lives on July 19, 1989, when they steered a McDonnell Douglas DC–10 with complete hydraulic failure to a crash landing at Sioux City, Iowa. That crew managed to (barely) control their stricken aircraft with an improvised method involving teamwork and differential thrust.

Goodrich said computers will need to be programmed to respond appropriately to such rare events, the “corner cases” where the plan gets tossed out the window in the face of the unforeseen, and that remains a challenge no computer has yet proved it can reliably handle.

Further complicating the task of automating the pilot’s job is the fact that present-day systems depend on human intervention, and if the human is “out of the loop” letting the autopilot fly, it takes time to develop full situational awareness in the event conditions present the computer with a math problem it cannot solve.

“We definitely have to change that paradigm of how the automation is designed and not expect the crew to be able to respond correctly 100 percent of the time in those corner cases,” Goodrich said. “So, the automation has to be made more capable, at least be able to stabilize the situation so that if there is still a single pilot on board, or if there are remote supervisors on the ground, they have the ability to come back into the loop more gradually, sort things out, before having to make some sort of dramatic change in the action of the flight, or the automation.”

Goodrich said machine learning today is generally designed to focus on fairly narrow and well-defined tasks, such as landing an airplane in predictable conditions, not so much on thinking to use the throttles to steer a DC–10 with no functioning flight controls. Bringing AI into parity with the human brain will require more refinement. Computers must be programmed to consider the river option.

“You have to be more rigorous in anticipating those challenging situations where today creative thought is required,” Goodrich said. “A more automated aviation environment would need to be redesigned around the strengths and also the limitations of what automation can and can’t do.”

Teamwork and trust

Automation advances have already improved aviation safety, Goodrich noted. Research into flight automation has produced, for example, systems in military use today that can step in if the human pilot loses consciousness during a high-G maneuver and pull the nose to level before the aircraft hits the ground. The automatic ground collision avoidance system credited with saving lives in just that scenario won the 2018 Robert J. Collier Trophy from the National Aeronautic Association.

But commanding the 5 p.m. daily flight to Cleveland, and being ready to handle whatever may arise, is another matter, and Goodrich noted that the time horizon for adoption of aviation technology—particularly when it comes to paying passengers—is measured in decades. Unmanned aircraft fielded by the U.S. military in the 1980s and 1990s are still not yet delivering packages routinely in the national airspace system. Synthetic vision and enhanced vision systems first demonstrated in simulation in the 1970s and 1980s are “just now coming online where you can get real, operational credit for them,” Goodrich said.

Another thing we’ve learned over the years is that the actual effects of automation on human jobs can be as difficult to predict as how many airplanes the world will need in 2039.

If Foley’s prediction that automation will enable single-pilot operation of transport aircraft in numbers by 2039 proves true, Goodrich noted, that might actually lead to a net increase in pilot jobs. Regional jets typically carry at least 70 passengers today, in large part because airlines have to pay pilots. The need to have a minimum number of paying customers limits the profitability of smaller aircraft flying shorter routes between smaller cities. So, one upshot of single-pilot airliners might be a proliferation of smaller aircraft serving destinations that airlines ignore today.

“Some pilot positions might go away, but other pilot positions could open up,” Goodrich said.

It often happens that way, historically, as industries are automated.

Goodrich does expect that aviation’s future will be more about evolving the present-day approach than a radical removal of human pilots from aircraft.

“The ultimate in performance and safety may be human-automation teams going forward,” Goodrich said. “Either type of agent has advantages and disadvantages that complement the other… My own bias is that human-automation teaming provides the greatest level of robustness.”

Aviation has for many decades required a very robust level of safety, with regulators and the public more intolerant of accidents in the air than on the roads. Smaller autonomous aircraft are likely to win the approval of both governments and paying customers sooner than larger aircraft will, and cargo haulers will be flown by computers before passenger carriers. Goodrich does not, by the way, expect increased automation can ever eliminate accidents. Computer-flown aircraft will be expected to crash, too, he said, though probably for different reasons.

Meanwhile, if you are starting or in the middle of your flying career, take heart, and continue to follow your passion, Goodrich advises. The job will probably change, and you may have a hard time making small talk with your artificial co-pilot in a few years, but aircraft will almost certainly have at least one seat for you, with flight controls, for decades to come.

“Aviation, in one form or another, is definitely going to involve skilled humans for at least out to that 20-year time horizon, and probably beyond,” Goodrich said.