Several systems able to control aircraft from before engine start until after shutdown have already been built and flown, though none have been certified or approved for use beyond constraints that keep them from mixing with other aircraft in the national airspace system, and the developers of these near-autonomy models are eager to start making cargo delivery flights that are widely seen as most likely to be approved by national regulators in the next few years.
Aircraft automation is one facet of a transformative change that is about to usher in new kinds of aircraft, in many cases powered by electric motors and designed to enable various new forms of “air mobility,” collectively termed advanced air mobility (AAM) by the FAA to include urban air mobility (UAM) operating from city rooftops, and regional air mobility (RAM) that would utilize larger aircraft operating from many more—perhaps eventually every one—of the roughly 5,000 public-use airports in the United States today. Making either of these new air mobility models work will depend in large part on more automated aircraft that can be repositioned on empty legs without a human aboard.
Avoiding risk to other aircraft or people on the ground that regulators and the public will not accept requires some combination of artificial intelligence and awareness of other objects and conditions. Various companies pursuing FAA approval as soon as late 2024 hope to facilitate that ambitious timeline by keeping trained pilots in the loop, monitoring flights remotely.
For example, if the Airbus A320 that US Airways Captain Chesley “Sully” Sullenberger ditched in the Hudson River in New York on January 15, 2009, had been automated in this fashion, nearly all of the tasks Sullenberger and first officer Jeffrey Skiles performed before and after striking the birds would have been handled by the automated system. Sullenberger would have made (or agreed to) the decision to ditch in the river, rather than attempt to reach one of the nearby runways, based on information provided that could include estimates of remaining glide range.
All contingencies, emergencies, and unexpected events will be handled in much the same way in any of the automation strategies currently being pursued.
Keeping humans in the loop will speed up airspace system entry and allow accumulation of flight hours and repetitions that will build trust that artificial pilots can rise to the kinds of occasions that have turned some pilots into folk heroes credited with “miracles.”
Hands (mostly) off
The last generation of pilots with stick-and-rudder skills may have already been born, though there are reasons to remain skeptical that all of this will actually happen. There is a great deal of money flowing into various efforts by government and industry to further develop aviation, and autopilots able to handle every phase of flight, including the decisions required to respond effectively and safely to whatever conditions or situations may arise, are central to making advanced air mobility economically feasible in the face of a chronic shortage of pilots needed to support current operations.
Pilots who have invested tens of thousands of their own dollars to earn the ratings, certificates, and experience required by their professional aviation careers will eventually be needed to help “train” their digital replacements, which could complicate matters.
When Airbus announced the first fully automated takeoff of a large passenger jet in January 2020, a video released with that announcement showed test pilot Yann Beaufils kept his left hand hovering over the sidestick after advancing the throttles. An artificial voice called out “V1” and the co-pilot followed with “rotate,” gently placing a hand on Beaufils’ arm as if to remind him to leave that rotation to the computer.
Such is the current state of trust in aircraft automation that pilots remain disinclined to let their hands stray too far from the controls.
Six months later, the company’s Autonomous Taxi, Take-Off and Landing project concluded after logging 500 flights, of which 450 were used to gather video to fine-tune the system that enables computer vision to navigate to and from the runway.
The airframer’s chief rival Boeing Co. is perhaps even deeper into the aviation automation development cycle, having gained years of experience and expertise with its 2017 acquisition of Aurora Flight Sciences, the Virginia company that has operated its optionally piloted Centaur, a modified Diamond DA42 piston twin, since 2012. Fully automated takeoffs and landings are old hat for Aurora’s Centaur, a general aviation aircraft able to operate autonomously (which strictly speaking means without human help) from before takeoff to after landing; it also has a hybrid mode that allows control from the ground with a safety pilot aboard “to facilitate testing.” And it recently graduated from test article to product when the Swiss military acquired in March the Centaur it had previously leased, planning to continue testing sensors and developing “a sense and avoid system for drones.”
In an email response to questions about the challenges posed by further development and eventual deployment of highly automated or autonomous aircraft, Aurora clarified that it does not expect to fully replace human pilots for some time.
“Autonomy technology is still years away from safe full autonomous operation in all conditions. Full autonomy will continue to be adopted where risk profiles are lowest (small UAS, remote operations, without passengers, etc.), but the short answer is human pilots remain essential,” the company wrote. “While it is impossible to accurately predict regulatory change and technical advancements, there is significant progress on the technology, testing, and infrastructure required to enable autonomous air mobility solutions, with many in the industry predicting timelines as short as five years.”
Beyond the world’s leading transport aircraft makers, well-funded startups have also flown test aircraft that are designed to handle all phases of normal flight without human help, though their leaders described in recent interviews an approach that incorporates human oversight of each flight.
In California’s Silicon Valley, Reliable Robotics and Xwing have attracted significant investment and forged partnerships with government and industry. Reliable Robotics announced in October a $100 million influx of venture capital that brought the firm’s total fundraising to more than $133 million. Xwing, having completed its first gate-to-gate automated flight test in February followed that with a $40 million funding round in April, and announced October 12 a partnership with Textron Aviation to jointly develop autonomous flight capability.
In both cases, Cessna Caravan variants are first in line for high-level automation. The aircraft have a long track record of safe operation; it is also the airframe of choice for many short-haul cargo operations. (Both of these companies were founded by and continue to be led by technology executives who learned to fly because they wanted to get places faster, and who remain frustrated that the transportation system as presently configured does not make that easy.)
Another U.S. company, Merlin Labs, is also led by a pilot, and is also retrofitting GA aircraft including Cessna Caravans and Beechcraft King Airs for military and civilian cargo service with its own near-autonomous flight control system, and announced in September that it has reached a certification basis agreement with aviation regulators in New Zealand for an automated aircraft that will operate under the oversight of an onboard safety pilot in order to build time and trust before taking the next step.
While Merlin’s machines will fly effectively “blind,” using ADS-B and the protections afforded by flying under instrument flight rules, the safety pilot aboard will provide the last line of defense against midair collision; Xwing has fitted its prototype with a combination of sensors to detect (and automate avoidance of) aircraft and other objects.
“It’s been one of the key missing pieces,” Xwing CEO Marc Piette said of the firm’s detect-and-avoid system that will develop artificial awareness of the aircraft’s physical environment using radar, lidar, and cameras in some combination, as Wired reported in 2020 after taking an uneventful ride on Xwing’s automated airplane.
Perfecting artificial situational awareness to the degree required for a highly automated (or fully autonomous) aircraft to steer clear of fuel trucks on the ground or other aircraft in flight, or to decide which distant strip of land is the best option for a forced landing may pose challenges, but perfecting such “awareness” might not be a prerequisite for gaining access to airspace outside of test ranges, at least on a limited basis—as long as humans remain in the loop.
“At Aurora, we work on both automation supported by remote pilots and on solutions that support single-pilot operation, whether that is automation, robotics, and/or autonomy,” Aurora Flight Sciences wrote in the emailed response to questions in October. “In many of the application areas we address, autonomy does not mean the absence of humans. Rather, it means decision-making for and with humans to perform in a trustworthy way.”
One way to help computers fly airplanes “in a trustworthy way” is to limit the potential loss of life that would result from an accident, and that is essentially why the industry aims to start with cargo flights on a limited number of routes.
“One advantage we have with automated systems is we get to pick where these things are going to fly,” said Reliable Robotics cofounder and CEO Robert Rose. Like Xwing, part of Rose’s plan to move the technology beyond test ranges and into the airspace system is to pick routes between airport pairs that have been studied and surveyed in detail. Along each route, potential emergency landing locations can be cataloged and assessed for possible future use. Failure modes and systems responses can be developed in simulation before they are ever tested in the real world, Rose added: “We actually have gone through this in the company … we have a very sophisticated simulation capability.”
Many possible emergencies such as engine failures can be planned for within such constraints, and preferable courses of action—including where the aircraft would land if the engine fails—can thus be programmed in advance. “You don’t have to figure this out on the fly.”
Aurora noted that machine learning will also play a role in preparing artificial aviators for the real world.
“AI agents can be trained from datasets, including previous incidents or near-misses, to learn a model that can anticipate conditions and quickly respond with [a] sequence of actions for safe continued operation. Imitation learning method can be leveraged to train the agent to perform a task from demonstration from human experts (i.e. expert pilots) by learning a mapping between observations and actions. This training methodology delivers both performance and understandability of how the agent is making its choices during the operation to foster trust with the humans in the ecosystem.”
Even with the most detailed planning and simulation, situations may still arise requiring a decision when the likely outcome or risk of adverse consequences is unclear, or when more than one option exists. “If it’s ambiguous,” Rose said, “that’s what we can present to the remote pilot.”
Rose said a similar process of planning responses to a vast number of contingencies was undertaken when he worked at SpaceX, designing the automated controls of the Dragon spacecraft that flew the first all-civilian crew into orbit and returned them safely home in September. While the spacecraft did give its crew the ability to take manual control, it was not needed or used.
While Xwing and Reliable Robotics are focused on equipping their aircraft control systems to handle the “aviate” and “navigate” domains, and Merlin Labs is devoting much of its programming effort to natural language interaction with other aircraft and air traffic control (the “communicate” piece), all three rely on humans to remain in a supervisory role, validating aeronautical decisions in real time, for years to come.
The human factor
Increasing aircraft automation begs another question: If human pilots are not actively involved in flying tasks beyond oversight and intervention only in rare events, will there be enough pilots willing to take this job, and will they be willing to help hasten their own obsolescence?
The prospect of being paid to see the world will no longer be part of the pitch if humans wind up overseeing flights from fixed locations. The “office” may not have much of a view, and stripping the last vestiges of romance out of aviation may impact pilot recruiting and retention, as it has in the case of U.S. Air Force remote pilots who report for duty in air-conditioned offices and fly aircraft from thousands of miles away.
The Air Force has struggled for years to meet its need for remote pilots, and the Government Accountability Office cited a few reasons in a 2020 report on the issue, including high workload stemming from a chronic remote pilot recruiting shortfall, limited opportunities for career advancement, and quality of life issues.
This also begs a related question: How many pilots, having invested years of life and tens of thousands of dollars achieving and maintaining commercial or airline transport pilot certificates and aircraft type ratings, will be eager to become remote operators of systems engineered to be as boring (a consequence of being “reliable”) as possible? That was central to one of four key challenges cited by McKinsey & Co. in a 2020 paper analyzing the costs and opportunities created by UAM in particular.
There will be pressure to move humans out of these cockpits, particularly the smaller aircraft that have few seats for paying customers. McKinsey authors Uri Pelli and Robin Riedel calculated that while the cost of operating these short-range, low-capacity aircraft roughly doubles on a passenger-mile basis in the presence of a human pilot, trained pilots somewhere in the control loop will be indispensable until regulators and the flying public accept and come to trust autonomous aircraft.
“These vehicles will eventually fly autonomously, but that could take a decade or more because of technology issues, regulatory concerns, and the need to gain public acceptance,” the authors wrote. “Another important challenge will involve creating a value proposition that will encourage people to embrace careers as UAM pilots despite the expense of basic flight training, the 12- to 24-month training period, and—most critically—an uncertain future. The UAM industry is quite vocal about the need to automate, potentially limiting the career of an UAM pilot to a few years. The net present value of a five-year UAM career could be quite low or even negative, given the upfront training cost and the opportunity cost of training time without income, even if compensation levels were in line with current early career pilots (around $40,000 to $60,000 per year). Further, UAM piloting skills and experience may not be transferrable either within or beyond the aviation industry. Many people might therefore believe it would be better to pursue other professions.”
It’s not clear how that will be worked out.
The FAA has not yet proposed any modification of existing pilot training requirements, and while the agency has worked with NASA and various industry partners to advance the development of simplified vehicle operations, establishing new standards and requirements for remote pilots to operate aircraft that require all of the decision-making expertise but none of the physical skills of current pilots, and finalizing any such rules, will take time. Part of that time will necessarily be spent gaining experience operating these aircraft remotely, and refining the process of making those aeronautical decisions from afar.
Pelli and Riedel wrote that such streamlined remote pilot training requirements “seem many years distant, however. Until then, prospective UAM pilots will have to take today’s training programs.”
While the pilot’s salary, training, and accommodations onboard the aircraft (seat and flight controls) account for around half of the operational cost McKinsey calculated in a UAM model, airline economics create less pressure to remove humans from the flight deck. Autonomous regional or long-haul air transport aircraft will arrive a long time after cargo operations begin, in part because the economic incentive to automate passenger service at scale is not nearly as strong, according to industry analyst Richard Aboulafia.
When the costs associated with human pilots are divided by more seats filled by paying customers, the passenger-mile cost of a pilot’s services drops quickly. While airlines spend billions of dollars per year on salaries, and these statistics are often cited by advocates of aircraft autonomy, Aboulafia said expenses related to employing human airline pilots—and therefore the potential savings that automation can achieve in the context of passenger transport in larger aircraft—are not hugely significant in the context of passenger air travel economics.
“Compared to fuel and capital costs and facility costs? No,” Aboulafia said. “Compared to passenger drinks and food? Yes.”
Airlines would have to automate much more than the flying task to realize significant reduction in their payrolls, dispensing with gate agents, flight attendants, ground handlers, or dispatchers, all of whom play a role in keeping the system operating safely and efficiently. The cost of pilots relative to overall airline payrolls “if you could isolate that out, and I’m not sure you could, it’s just not that much.”
Aboulafia is skeptical that automation will extend beyond a few limited use cases, such as cargo flights by small aircraft, for decades to come, possibly a half-century or more.
“It’s anyone’s guess, but that seems reasonable,” Aboulafia said. “I don’t even know if it works out ever. If you look at the cost structure associated with regional (airline) operating economics, they’ve been pretty bad for some time now. Getting rid of the pilot really doesn’t do much.”