Sometime next year Mike Sinnett, Boeing’s vice-president of product development, will enter a small, experimental aircraft and – he hopes – do nothing.
Boeing Commercial Airplanes (BCA) has started exploring autonomous flight technology for passenger-carrying aircraft and Sinnett, as a pilot and engineer, plans to fly in – as opposed to “fly” – the first test subject. Boeing’s newly developed, machine-learning software is already loaded into a flight simulator, which Sinnett and his team have been using to refine the algorithms. But the real test will come next year when flights begin with Sinnett on board, as the software makes decisions that respond to changes in the environment.
“I’m not ready to talk yet about what those decisions are,” Sinnett says, speaking to journalists earlier this summer. “And I’m not going to close the loop on the airplane. But I’m going to make sure the decision is made with the same set of inputs that pilots use to make decisions and I’ll record the decision that the airplane makes.”
Boeing has not publicly identified the aircraft Sinnett will use next year, except to describe it as a small and far less complex than a commercial transport. But the size and complexity of the test aircraft will escalate several levels in 2019, as Boeing reintroduces a 787 into the ecoDemonstrator fleet. Taking incremental steps towards greater autonomy, the ecoDemonstrator 787 will incorporate software to manage taxi and take-off in place of a pilot, Sinnett says.
The ecoDemonstrator is tasked with evaluating technologies that could be used on future or existing Boeing aircraft. By studying new autonomous control modes on the 787 ecoDemonstrator in 2019, Boeing could have the technology ready to appear on its next clean-sheet aircraft. Boeing has proposed developing a family of new aircraft after 2024 to fill a perceived gap between the 737 Max 10 and the 787-8.
It’s an extraordinary move within BCA. Although sister businesses in the defence and space markets are deeply experienced with autonomous vehicle control, BCA’s approach to cockpit architecture for passenger-carry transports emphasises that a human pilot has ultimate control. Even in an age with flight envelope protections enabled by fly-by-wire controls and auto-landing systems, the pilots of 777s and 787s are never “out of the loop”.
Sinnett acknowledges the cultural shift, then points out that BCA is not yet committed to developing an autonomous airliner. “We're not going there yet. We're exploring,” he says.
Indeed, there are few signs the commercial transport market is prepared for such a disruptive shift. As a whole, the industry is more profitable than ever and all projections point to continued traffic growth for the foreseeable future. But that very traffic growth sets the industry up for a tough challenge: where are all the pilots going to come from? To meet projected demand for new aircraft, Boeing estimates that airlines will need to hire two million workers over the next 20 years, including 637,000 pilots. An ever-shrinking pool of military-trained pilots means airlines could struggle to find enough classic “aviators” with a rich depth of aviation knowledge and expertise.
Twenty years from now, Sinnett wonders, are pilots going to be operators of machines rather than aviators? “That drives you to think of things differently,” he says. “The pilot is ultimate authority of a commercial aircraft today, but that’s an experienced pilot with the right level of proficiency and the right level of aeronautical knowledge. If the assumption that all of those pilots will always be available is shown to be an invalid assumption 10 years from now or 20 years from now, then we have to have a different plan.”
Boeing is not the only company contemplating the possibility of passenger-carrying aircraft with fewer or no flightcrew within 20 years.
The research arm of Swiss bank UBS published a report on 7 August that notes it would be feasible to operate “remotely controlled airplanes carrying passengers and cargo” by about 2025, potentially saving the world’s airlines $26 billion a year in foregone pilot salaries, reduced fuel bills and lower training costs. Although bank’s researchers elaborated on the financial benefits of a shift to pilotless aircraft, they recognised that the industry was unlikely to be ready within eight years to employ such technology. And, even if the industry can overcome regulatory barriers, airlines can expect to find that the population is mostly unwilling to fly in a pilotless aircraft.
Still, if it can be achieved, an automated cockpit solves two of the industry’s most intractable problems at the same time: a pilot shortage and creeping labour costs.
The regulatory barriers, however, are significant. For this reason, the industry seldom transitions to the full employment of a new technology in one great leap. There are usually smaller steps. A common example is the transition to carbonfibre-based structures instead of metal. The first applications appeared in the 1970s on secondary structures, such as the rudder for the Airbus A310. By the early 1990s, Boeing was ready to replace metal with carbonfibre on the empennage of the 777. More than 15 years later, the 787 entered service with a carbonfibre fuselage and wing – nearly 40 years after the first application.
Some would argue the transition to an automated cockpit also started about 40 years ago. That was when US and European regulators accepted a two-person cockpit, which removed a requirement for a navigator. Since then, the industry has introduced new autopilot features, including autoland, which allows the aircraft to navigate final approach and landing by itself in certain situations at qualified airports.
Sinnett offers a possible roadmap for a step-by-step, incremental transition from the crewed cockpit of today to a fully pilotless aircraft. He notes that some airlines that operate a 777 on a 16h mission require five pilots on board: a captain, a first officer, a two-person reserve crew and one pilot dedicated to the cruise stage of the flight. By introducing more automated redundancy in the cockpit, the five-pilot crew might be the first thing to go.
“Some of the first steps might be to go from five [pilots] to four, and then from four to two to reduce the number of augmented crewmembers on the flight. That may be the first step along the way,” Sinnett says.
“Another step may be to go from two pilots during cruise to one pilot during cruise and [another] pilot on board the airplane, but maybe getting meaningful rest. It could be that you have a one-pilot operation.”
Single-pilot cockpits are banned for most types of commercial operations today, but there are exceptions. Sinnett notes that the US Federal Aviation Administration allows certain airlines to fly up to 10 passengers with a single crew member. One example is US regional carrier Cape Air, which operates nine-seat Cessna 402s with a single pilot.
“We as a society are willing to accept the risk – given the size of the airplane, the number of people on board and the weight of the airplane – that it can be operated by a single pilot,” Sinnett says. “As a society you can ask the question: if it’s okay for a single pilot to fly 10 passengers in a certain airplane type, why would it not be okay for a single pilot to fly a freighter with no passengers on board, and right now that is not allowed. That is also potentially one of the steps along the way.”
Of course, the step beyond single-pilot is no pilot. As Boeing considers the path for introducing higher levels of automation, the company still is not sure whether this should be the last step or the first. In the latter example, the industry would bypass the step-by-step process and leap as quickly as possible to a pilotless cockpit.
“What isn’t clear yet to anyone in the industry – ourselves included – is whether it’s a single step from what we have today to full autonomy, or whether it happens in step-wise improvements over time – each of which retains the same level of safety integrity that we have today. We don’t know the answer to that question,” Sinnett says.
“You can imagine if you took those successive steps it might take a lot longer to go from where we are today to all the way. You can imagine six steps to autonomy – each of which would be very, very difficult, each of which would be a battle in its own right. So maybe taking each step isn’t the right answer, and that’s part of what we’re trying to figure out.”
The critical challenge is meeting the industry’s standards for safety. Driverless cars are quickly becoming a reality, but the US automotive industry faces a different bar for safety. In 2016, for example, more than 40,000 Americans died on roads, but none died on airlines in US airspace.
“So that drives a very different way of thinking about the problem,” Sinnett says. “We have to have the same level of integrity that we have today.”
Aircraft already possess multiple automated functions, which Sinnett lists: autopilot, autoland, autothrust management, auto-navigation, aircraft health monitoring and reporting. These systems are automatic but not autonomous. At least two pilots are on board and assigned to monitor each function and intervene if anything goes wrong.
For example, the autopilot fails in very rare cases, Sinnett says. Suppose the crew has programmed the autopilot to make a turn, but then it doesn’t and the aircraft continues flying in a straight line. The pilots are on board to recognise such problems, he says. They would disconnect the autopilot, make the turn manually, then reconnect the autopilot while making a note to report the incident.
Such a scenario involves a functional failure but not a safety issue, Sinnett says, since a human intervened to solve the problem The system is designed to be extremely safe, with any quirks managed with human monitoring and intervention. In a fully autonomous aircraft, the systems would have to be reliable enough to manage themselves.
“Without the pilot in the loop to catch that first link, it begs the question, what would the next thing be that happens? Would the airplane go five miles off course?” Sinnett asks. “Some of the work we’re doing today is to try to figure out where all those gaps are in the design of an airplane and how you would close those gaps successively through a series of steps that go from where we are today to full autonomous operation.”
Boeing may also have to persuade regulators not only to accept autonomous systems, but to change the way they verify that software is safe today. The most advanced software in aircraft today is certificated as airworthy using a prescriptive series of tests. To be certificated, software code is given a set of inputs and it must generate the same set of outputs without variation. A fully autonomous system, however, uses machine learning software, which reacts differently to situations as flight conditions changes – sometimes in ways that are impossible to anticipate.
“Nobody is smart enough to program all the potential things that can happen in the operation of the airplane and then demonstrate the airplane does the right thing all the time. So we have to come up with a different way to do it,” Sinnett says.
Europe’s approach to autonomy
While Boeing has announced its ambition to develop autonomous (pilotless) capability for commercial transport aircraft, Europe has hedged its bets.
EASA admits a relaxed approach to companies developing pilotless small aircraft – potentially including air taxis – and the European Commission (EC) has financed an industry-wide research programme called ACROSS that aims to enable both reduced crewing and increased safety in large commercial aircraft.
The objectives of ACROSS – advanced cockpit or reduction of stress and workload – are to enable the development of technologies to help pilots in a number of ways. These are: to cope with peak workload (dense traffic, bad weather, emergencies); to enable reduced-crew operations, including single-pilot crews, through the use of a crew monitoring system; and to cope with crew incapacitation by providing an “electronic standby pilot” that would recover the aircraft to the nearest suitable airport.
Participants include big European companies such as Airbus, Thales and BAE Systems, along with smaller companies, universities and interest groups such as pilot organisations.
The programme concluded and published its findings last year. Solutions identified by ACROSS, according to the EC, are designed to help industry develop “useful tools, technologies and guidelines” that will enable progress towards autonomy for large commercial transport aircraft, while addressing the pilot supply problem by enabling reduced crewing.
These solutions will include: technology for automatic crew monitoring; new avionic functions to improve the performance of all basic crew tasks (“aviate, navigate, communicate and manage systems”), especially during peak workload situations; and the technical capability for continued safe flight and landing in case of crew incapacitation. These technologies, plus specialist crew training, could enable the safe management of reduced crew operations.
The ACROSS summary calls for development of advanced displays, capable of providing guidance for upset recovery along with aeronautical information and mission management assistance. It also calls for advances in controls and interaction, and in automation and assistance. Better crew monitoring systems are also on the agenda.
Meanwhile the principal adviser to the flight standards director at EASA, Yves Morier, is relaxed about companies developing autonomous small commercial aircraft such as air taxis. The way the regulators think about airworthiness and certification nowadays is not as prescriptive as it used to be, he says. To paraphrase the EASA attitude: “Bring it to us with the test results, and if you can prove equivalent or improved safety compared with piloted vehicles of the same size and in the same role, it will be approved.”
In fact, when true autonomy becomes sufficiently sophisticated to entrust it with the control and management of widebody passenger aircraft, Morier suggests, it will be through lessons learned at the smaller end of the marketplace.
How will regulators approach pilotless aircraft certification?
At present airworthiness regulations in all countries require at least one qualified pilot to crew a commercial transport aircraft. The aircraft commander has legal responsibility for its safety and signs documents to that effect before every flight, and in airworthiness terms he or she is an essential component of the aircraft. The pilot’s ability to make decisions and, when necessary, to manipulate the aircraft and its systems is an assumed part of the total system from the time the aircraft’s designers begin work on it.
Yves Morier, the principal adviser to the flight standards director at the European Aviation Safety Agency, points out that since a pilot is an essential design component in a commercial air transport aircraft, in the case of a pilotless commercial aircraft every legal responsibility the pilot takes on, and every task the pilot is expected to be able to perform, has to be achieved by alternative means.
Yet Morier says he can see small pilotless commercial transport aircraft – perhaps operated by an Uber-type air taxi company – possibly certificated within about 10 years. He adds that these will, in early days, almost certainly be limited in what size they are, what they can do under what circumstances, and in what kind of airspace environment they can operate.
The unspoken implication is that certification for larger commercial aircraft is a rather more distant prospect, but experience gained from the bottom of the air transport market about the reliability of the enabling artificial intelligence (AI) systems controlling these vehicles in the aviation environment will advance the application of AI in more sophisticated large transport aircraft. And, because the essential characteristic of AI is the ability of the computer to learn and apply its learning, the systems may develop credible capability at rates the world is not accustomed to.
Morier says developers are looking at pilotless air taxi prototypes flying in the next couple of years with planned service readiness dates of 2023-25. From his prediction of a 10-year wait for approval, however, it would seem that the regulators are rather more sceptical.
Meanwhile unmanned aerial vehicles, whether military or civil, are becoming a familiar fact of life, particularly as small drones become ubiquitous. Obviously, then, the technology to enable simple pilotless aircraft to operate already exists, but using the same relatively unsophisticated technology for the task of commercial air transport, and operating such vehicles in the same airspace that airlines use, is unthinkable. The non-combat UAV attrition rate, even of sophisticated military drones, at up to 30% makes that clear.
Fourth-generation airliners are highly automated, and already fly almost all the flight profile from take-off to landing without physical intervention by the pilots. Yet US studies have shown that less than 10% of all flights are performed exactly as flight-planned, and the reasons for that include factors as diverse as weather changes, air traffic control and traffic considerations, technical malfunctions, and emergencies on the ground or in the aircraft.
With present flight management systems, one circumstance guaranteed to make them go off line is failure or malfunction of some of their sensors. If they cannot get information, or if data inputs conflict because of malfunction, the system is programmed to recognise that it cannot function usefully, and to disconnect, handing the aircraft back to the pilots.
Vital basic functions still performed by the pilots include simple acts such as selecting the landing gear down or up, and likewise the flaps. These activations would be simple to automate, but the recognition of when they should be carried out – and when they should not – is more complex.
AI will have to learn to cope with “Black Swan” occurrences, Morier observes. This is the industry expression used to describe combinations of circumstances that were not foreseen and for which there is no procedure. And since, by definition, Black Swan events may not ever have occurred before in precisely the same way, and may never happen again, the opportunities for AI to “learn” about them is compromised. This would include combinations of failed or conflicting data inputs.
Even pilots, however, cannot be guaranteed to manage Black Swan events well, but those who do tend to cope by prioritising the essentials according to the traditional mantra: aviate, navigate, communicate. It would surely not be beyond the wit of man to program similar priorities into an avionics system, but the problem might be to get it to recognise what to ignore when reverting to basics, because in the cacophony of failures that can follow an emergency such as an uncontained engine failure that also damages the airframe, the computer’s reaction might not be predictable.
In a different era, but not long ago, it took decades to prove to the regulators that engines were becoming sufficiently reliable that just two would be accepted to power commercial flights over the North Pole, northern Siberia and the big oceans. More recently, it took from 1992 to 2014 for Europe to agree conditions for permitting commercial passenger operations in single-turbine aircraft in instrument meteorological conditions and at night.
But they are accepted now. Morier says a different mindset is required by regulators today, because technology is moving so fast.