Flying Smarter: AI & Machine Learning in Aviation Autopilot Systems
Many experienced airline travelers have experienced a common scenario- As the boarding time approaches, the gate agent announces that there will be a slight delay since the pilots are running behind schedule on their current flight. The passengers continue to wait as the plane and flight attendants sit idle on the tarmac. Several companies are working to transform the cockpit of the future to reduce or potentially eliminate the requirement for manned pilots, which would make this experience a relic of the past.
The average passenger is often unfamiliar with the level of autonomy currently provided by the autopilot systems in use today in commercial aviation. These systems have been around in various forms since the 1980’s. For commercial flights on a modern aircraft with greater than 18 passenger seats, the airplanes flight management system (FMS) and its associated autopilot functions are generally in control of the aircraft from shortly after takeoff until landing and rollout on the runway under normal operations. Instead of flying the aircraft manually through the flight controls, the crew manages the aircraft’s systems through the FMS interface. Nearly all major airports utilize the CAT IIIb “autoland” instrument landing system approaches where the pilot is only a backup to the automated landing system in the event of a failure and takes over once the plane is on the ground to taxi to the gate. CAT IIIc approaches that would extend this autonomous control through the taxi process are in development in many places as well. These automated systems do have their limitations. In the event of a mechanical issue or extreme environmental issues such as severe turbulence, the autopilot may disengage and alert the pilot to take manual control of the aircraft. The system is also only as “smart” as the pilot that is inputting the information, and they require careful attention and oversight to ensure that the system is functioning as required. Many high profile incidents, including the tragic Air France Flight 447 crash in 2009, have resulted from miscommunications between the aircrew and the autopilot system. The pilot is still ultimately responsible for flying the aircraft, and these systems function as a kind of advanced aerial cruise control that simplifies and in many cases eliminates the hands-on flight control manipulation to allow the pilots to focus on communicating, navigating, and managing the flight as a whole.
Several groups are working to transition from the current generation of autopilot to an artificial intelligence and machine learning driven autonomous or semiautonomous aviation future. Current systems are hard coded to deal with a limited range of input information consistent with “normal” operations, and the systems are designed to cede control to the pilot for any situation outside these parameters. A team of AI experts from the University College London have researched applications for machine learning algorithms to enable a next generation autopilot system to learn to handle unexpected situations by feeding the computer the responses of trained pilots to similar scenarios in a flight simulator. Over time, the system has demonstrated the ability to respond to engine failures, turbulence, and extreme weather to maintain a level flight profile. The system also shown the ability to fly simulated aircraft with different design characteristics than the model was trained out, demonstrating a great deal of flexibility. While the implementation on an actual aircraft would require a much larger neural network than the network currently required in the simulator, the team’s research points to a promising future for the technology. Several large airplane manufacturers including Boeing and Airbus are working on similar research.
A key hurdle to overcome is that the functioning of a neural network of the size required to fly an actual airliner would amount to an intractable black box that will make testing and certification by aviation regulatory bodies difficult. Despite the fact that human error accounts for over 80% of modern aircraft incidents, airline travel is the safest it has ever been. Regulators will require a high threshold of proof that these systems are safe and effective before allowing them to transport passengers. Still, the implementation of this next generation of technology will allow airlines to further reduce the number of pilots required from two or three down to a single pilot, reducing the impact of the looming pilot shortage that is currently forecast. These more advanced systems will help alleviate many of the issues that currently arise at the boundary between the current autopilot systems and their manned counterparts, ultimately improving the safety and efficiency of air travel.
