Generative AI for aeronautics and outer space applications

Mykel Kochenderfer is Associate Professor of Aeronautics and Astronautics at Stanford University. Prior to joining the faculty, he was at MIT Lincoln Laboratory where he worked on airspace modeling and aircraft collision avoidance, with his early work leading to the establishment of the ACAS X program.

We should first recognize that aviation as a sector is a miracle — flying in a metal tube comfortably at 35,000 feet at a speed of 900 km/hr. To get us to our destination safely, we needed collaboration between education, investment, manufacturing, services and regulatory sectors. Since society has been enjoying decades of safe travel, the natural course when a new development is on the horizon is risk aversion, which is perfectly understandable. Regulators tend to be hesitant to adopt new technologies, even if they enable new applications and improve safety.

Aviation as a sector is good at learning from mistakes and in sharing data. Safe travel requires a culture of safety and oversight of manufacturing, design, training of pilots, maintenance and repair. IATA for example launched its ‘Turbulence Aware’ initiative in 2018 to help airlines mitigate the impact of turbulence, the primary cause of passenger and crew injuries in the air. The program currently has more than 20 airlines feeding real-time data into the system. IATA has a goal of collecting turbulence reports from 150 million flights by the end of 2024. Airlines expressed a greater interest in the program after the Singapore Airlines incident early this year that left one dead and scores injured.

(Infosys Data + AI Radar reported that generating real value from data and AI requires companies to share data. Most organizations have learnt how to share data tactically. Strategic data-sharing practices allow an organization to share data quickly, with control for specific opportunities with trusted partners, and to share it repeatedly, with an emphasis on value creation.)

We should preserve a culture of safety and a certain degree of risk aversion, but fund initiatives to address challenges in civil aviation. We should broaden the expertise available today, with the combination of multiple domains. Traditional engineering and education have been siloed in nature. Systems and sub-systems in an aircraft continue to reflect this in the form of challenges to integrate these systems to make them work as a coherent system.

Language and visual encoding for flying

Autopilots or automating a pilot’s tasks started as early as 10 years after the Wright brothers flew their first flight in 1903. Kalman filters have been used for navigation for decades since the 1960s. Subsystems have historically been automated. However, the interface between systems is multi-disciplinary.

Aircrafts consist of diverse systems such as propulsion, avionics for navigation and communication, flight control, environmental, and safety systems. It is a challenge to integrate these systems to work as a cohesive whole due to the need for common standards, communication protocols, interfaces, and architectures. A lack of interoperability can hinder communication and coordination among the systems. There are trade-offs between performance, reliability, safety, and security.

Generative AI can be used to spur creative approaches to the design of future aircraft. It may have the potential to develop novel aircraft structures with better aerodynamic properties and performance profiles. It can lead to advances in material science with lightweight, composite materials. It may even help with the design of sustainable aviation fuels.

The recent developments in generative AI support the potential for autonomous flights in a few ways. Two broad approaches are the language side, and the visual side. The benefit of language models to the aviation industry may seem surprising. There are concerns, of course, about whether these language models will hallucinate and cause issues. However, large language models (LLMs) can encode a huge amount of common sense. A pilot’s handbook, emergency procedures and accident debriefs can be incorporated into these models.

On the visual side, transponders and sensors have been traditionally used in aviation for navigation, together with ground-based radars. We have invested in infrastructure with flight navigation tied to the capability of radars for more than 50 years. AI might be able to provide a visual understanding in situations where collision avoidance and navigation must be done by visual input. In the case of eVTOLs (electric vertical take-off and landing vehicles), visual guidance could be useful in situations such as flying between buildings or to manage a potential bird hit.

Generative AI in outer space

Space exploration is a multidisciplinary domain where errors are costly, and their impact magnified many times if there is human fatality. Robotic planetary exploration must be robust to long communication blackout periods and time lags. AI can help such systems manage themselves autonomously.

AI can help build highly reliable autonomous systems that take into account various sources of uncertainty. AI can also synthesize test cases and validate them beyond the capability of humans. However, one big obstacle for deploying AI is the validation process and establishing confidence that the system will operate as desired.

Stanford HAI and Infosys

The mission of Stanford University’s Human-Centered Artificial Intelligence (HAI) is to advance AI research, education, policy and practice to improve the human condition. Infosys recently announced its collaboration with Stanford HAI. As a member of Stanford HAI’s Corporate Affiliate Program, Infosys will have the opportunity to leverage and contribute towards leading global AI research initiatives. Application of AI in the aerospace sector can be an area of collaboration between Stanford HAI and Infosys.