Future of Flight: Autonomous Aircraft?
How can you reduce the likelihood of mid-air aircraft collisions? Can the solution be a pre-cursor to fully autonomous flight? This article explores the use of machine learning to replace traditional Traffic Collision Avoidance Systems (TCAS).
You are driving down the street. You have been down this same road almost every day for the past decade. Home is just past the next intersection. As you regain speed after halting at the ‘Stop’ sign, a driver screeches to a stop just inches from your door. He had not planned for a full stop at the sign. Adrenaline rushes through you, but everyone is okay. So, you head home. You resolve to ask for a traffic light at the intersection during the next community meeting.
Pilots sometimes find themselves in similar situations. They missed a command from Air Traffic Control (ATC). Another aircraft lost its radio ability. The instruments provided a bad signal. Even if a collision is imminent, there is no way to come to a full stop mid-flight. The pilot must be aware of potential risks and plan for them.
The Grand Canyon mid-air collision in 1956 spurred the Federal Aviation Administration (FAA) into action. The current solution, Traffic Collision Avoidance System (TCAS), was first implemented 25 years later. Required on all aircraft weighing over 12,600lbs or carrying more than 19 passengers, this system works by constantly broadcasting each aircraft’s altitude and bearing. [1] When a nearby aircraft is detected, it announces a verbal instruction to the pilot to ascend or descend. TCAS instruction take priority over everything else, including ATC; however, it depends on the pilot responding in time and correctly. Still, the pilot may not be able to comply: the aircraft is too close to the ground, its aerodynamic capabilities prevent action, or the fuel levels are too low. Even so, TCAS has proven its usefulness many times over.
The 2001 Japan Airlines incident [2] and the 2002 Uberlingen collision [3] proved that TCAS is far from perfect. Automatic Dependent Surveillance – Broadcast (ADS-B) is a revolutionary technology set for implementation in the US airspace starting in 2020. ADS-B accounts for a much larger set of inputs – type of aircraft, runway occupancy, weather, etc. – and provides much more comprehensive guidance to pilots in at-risk situations. [4]
The complexity of correctly matching an output given the expanded number of inputs has necessitated the use of machine learning. Using clustering, ADS-B transforms scattered flight data points into individual continuous flight trajectories. Learning by observing actual pilot action under different input conditions, the system then teaches how to best respond to a wide range of situations. [5] It would not have been possible for the system to reach the same conclusion using traditional solver algorithms in a reasonable amount of time (NP-Hard). Millions of flight hours have already been logged to train ADS-B systems.
ADS-B is set to be included on a much larger set of aircraft, including those that fly Visual Flight Rules (VFR). Longer term, there are proposals to expand the system to fuel trucks, baggage handlers and other moving pieces involved in aircraft operations. [4]
Air travel has become ubiquitous globally. Despite better pilot training and cutting-edge technologies, the increase in complexity of airspace management has outpaced this progress. Collaboration will be key to maximizing the effectiveness and benefits of ADS-B.
To engage research scientists and stakeholders (pilots, ATC, etc.), the FAA needs to find ways to balance ADS-B data security concerns with the added learning that can be gained. Like how Uber uses data to optimize their products, ADS-B data can be used to proactively predict and prescribe improvements to flight patterns and ground movement at airports. Such optimizations can ensure that the occurrence rates of high-risk situations under normal operating conditions is close to zero. Under stressed conditions, this learning can inform where the key players should focus their attention.
The density of and variability in air traffic is going to increase. ADS-B can help optimize the global airspace in addition to its primary safety benefits. The data can be used to understand where margin can be repurposed and where more is needed; machine learning can ensure that FAA’s ability to meet customer needs is maximized while holding extremely high safety standards. Additionally, such systems can help prepare for the introduction of drones to commercial airspace.
As with autonomous vehicles, an adequately large data set provided to machine learning algorithms will push the expected safety levels from such systems to beyond those of human pilots. Would you trust your aircraft if it was fully autonomous? Longer term, would you be comfortable with decoupling the pilot from the aircraft i.e. ground-based pilots are available to take over cockpit-less aircraft when needed?
A similar system has been proposed for ground-based autonomous vehicles. Each car would broadcast its details, reducing the load on other cars to determine their surroundings. Do you see value in integrating such systems with ADS-B? How would you balance the benefits here with the added complexity and implementation delay it would result in?
(795 words)
Bibliography
- Federal Aviation Administration. 2011. “Introduction to TCAS II.” https://www.faa.gov/documentLibrary/media/Advisory_Circular/TCAS%20II%20V7.1%20Intro%20booklet.pdf
- Flight Safety Digest, March 2004
- News, ABC. 2018. “52 Kids Among Dead In Midair Collision”. ABC News. https://abcnews.go.com/International/story?id=79916&page=1.
- “Automatic Dependent Surveillance-Broadcast (ADS-B)”. 2018. Faa.Gov. https://www.faa.gov/nextgen/programs/adsb/.
- Sun, Junzi. 2016. “Large-Scale Flight Phase Identification From ADS-B Data Using Machine Learning Methods”. Presentation, TUDelft, , 2016.
Great article and I enjoyed the preview of the future. My main concerns with ADS-B system are not the incredible efficiencies or whether the technology should be implemented. I know that computers often perform better than humans in certain tasks and flying would appear to be one of them. However, I’m more concerned about the effects of a system outage on the larger network. Today, each pilot’s actions within their plane is independent (for the most part). But once all flight paths are acting on the same technology, we’ve created an intricate network that is susceptible to Chaos. How do we create protections within the system to isolate the effect of outages and maybe even more importantly, protect against targeted attacks.
Great article! I personally dont believe that ADS-B will ever allow for a pilot to be decoupled from the aircraft because the aircraft is holding human lives. All it would take is for one malfunction of ADS-B to decimate the autonomous flying industry even if a remote pilot could be relied on to steer the plane to safety. Although, if a remote pilot where necessitated, one could argue how much thought process would go into the decisions he is making because his life doesnt depend on it. Therefore, I believe the machine learning in flights, or ADS-B, adds capabilities to the pilots, but it cannot replace pilots.
Excellent post. I wonder however what are the risks of a malfunctioning system and whether I trust it over a human pilot. Afterall, statistically speaking the number of flight collisions compared to the total global flying hours remain very small. Also, is the cost if the system low enough to warrant the investment? I am of the opinion that ADS-B will be a nice to have but not yet a necessity unless other features are added.
Great post! It is intriguing to see how machine learning can help solve a life threatening problem and actually ensure better decision making and control over the aircraft. I am more intrigued to find out how the number of accidents can be foretold using the data that is collected using the machine learning algorithm. Since this is a use case where even one mistake could mean loss of life, the accuracy of the tool is paramount. How d we ensure that is accuracy is achieved?
This was such an interesting article, thanks for your insight! I had not thought about the complexities of flight outside of controlling the actual aircraft and the complexities of taking off and landing. I think that until we see full implementation and success of automated cars, society will not accept ADS-B decoupling pilots from the aircraft. Although flying is statistically safer than driving, people still have an irrational fear of flying that would absolutely amplify if machine learning replaced human control. However, people are already aware of aircraft using autopilot to help fly, so if this system were explained in the same way it could have a greater chance of success.
Great article, thanks for sharing these interesting insights. I am convinced about the benefits brought by ADS-B systems in terms of safety. The applications towards autonomous flight seem less realistic to me. While I believe that advanced AI can compute an optimal trajectory to avoid collision, or to plan for take off or landing (which are the main critical phases that still need to be done manually by a pilot), I am skeptical that a machine could act on the plane controls to follow that trajectory. Mostly because it is not just about the speed and steering wheel like in an autonomous car. In 3D (as opposed to cars) there are uncertainties related to turbulences, inclination of the plane, etc. And as you point out, machines are only trained with “successful scenarios” so they don’t really know what a mistake look like, or how to deal with an unexpected response from the plane. Would it be possible to train the machines with real pilots, on simulators, to improve the level of training?
Finally, I would be curious to see how significant the impact on plane manufacturers would be (impact on design especially) and if autonomous planes would impose particular constraints on airports, airlines and other stakeholders, especially in terms of accountability.
Thanks Akash for sharing your thoughts! Truly enjoyed learning about recent developments to embed the use of machine learning in air traffic control. My guess would be that this is a pretty under-innovated space (e.g. I’d assume the FAA servers run COBOL – or something close to it.
View views are both optimistic on the need, but bearish on the solution. Air traffic control and autonomous guidance of aircraft is going to continue to become more important with the introduction of more drones (today) and personal multi-copter aircraft (near future). However, my concern around using machine learning as a panacea is that it’s a “black box” – pun intended. Especially with complex models, while the behavior may be “smart” the results and how it gets there are actually not interpretable. So in a highly regulated context, when you cannot explain a “this = that” relationship, how can policy makers and engineers evaluate deployed solutions? Maybe we have to settle for basic stuff built robustly…
Thanks for the insightful article and great insight. I too worry about the long-term implementation of ADS-B to further take tasks from human pilots, particularly in commercial aviation. I think it comes down to the question about trust that you pose in the second to last paragraph – that is, that consumers would be comfortable sitting in an autonomous aircraft. Given the significant PR issues surrounding recent incidents in autonomous driving – which have brought with them stringent regulations and public unease, one such issue in air travel could render this technology obsolete for years to come. Regulators would be very careful about approving this project, if at all.
Great article! I personally would always want to have a human pilot at least available to take over at all times. But I also would not want a pilot without ADS-B available either. So I think rather than a substitute for pilots i view ADS-B as a complement and do not see fully autonomous aircraft becoming widespread in the near-term even if the technology has been developed.
Thanks Akash, super interesting! I think ADS-B is a great advancement that has a lot of promise for improving the safety of flights. Similar to our Watson conversation, I think it’s a great supplement to pilots -to help inform them of what to do and how to react to certain alerts and scenarios. However, I think that as we develop this tool and knowing machine learning faults, we are bound to discover limitations of the software, exceptions, and instances where the software might have shortcomings and we might need a human to step in. Given the highly binary outcomes of airplane accidents, this is a high stakes situation and I would be nervous to have a remote pilot (what if the system goes down or gets hacked, etc.) Therefore, I think this is a tool for pilots rather than a complete substitute!
Very interesting article! Personally I would not trust an aircraft if it was fully autonomous neither if the pilot was not on-board. For two reasons:
First, I think machine learning has great potential and should definitely be used in systems such as the one described by Akash. However, machine learning depends on training with past data. When facing new, unexpected events, the creativity and ingenuity of a pilot may be a critical competence to prevent accidents. I believe the automatic pilot should control the aircraft during most of the time, but we need a pilot to take over the control if such new, unpredictable events happen.
Second, although I believe the pilot may effectively control the aircraft remotely, I would be concerned about situations in which some unexpected events harm the communication between the pilot and the aircraft. Besides, and maybe a controversial topic, I would also be concerned to whether the pilots would have their full attention and would be performing at their full potential if they are actually not in the aircraft. I’m sure they have the best intentions, but placing the pilot in the aircraft itself is the best way to make him fully committed to prevent accidents, as his life is at stake.
Thanks Akash! A future world of ADS-B enabled commercial flight is superbly interesting, awesome promising, and slightly scary. One of my key concerns is on the system’s susceptibility to hacking and other nefarious activity, I concern I also share for ground-based automated vehicles, and even airport air-trains! Government electronic security agencies routinely warn about the vulnerability of critical infrastructure, like power grids, for example. I wonder what sorts of controls need to be put in place to minimize these risks.