Fabrice Valentin is VP at Airbus headquarters in charge of the group’s artificial intelligence strategy. On the occasion of VivaTech, he returns for JDN to Airbus’ trajectory in terms of AI, from its experimental DragonFly project to the recent partnership with Mistral AI.
JDN. Where is the DragonFly project today, which aims to provide aircraft with autonomous capabilities at Airbus?
Fabrice Valentin. A first phase of this project was completed a few years ago under the aegis of UpNext, our innovation subsidiary. But DragonFly made babies. A large part of the demonstrations revolve around the visual landing. It’s a bit like the Holy Grail: as soon as we associate AI with Airbuswe immediately imagine that we are going to automate planes. As head of AI across all of our areas, I remind you that this remains research and long-term. On the other hand, what DragonFly demonstrated is that AI must be confronted with the real world. It must be practiced on concrete cases to validate the capture and quality of the data, and ensure that the system works. Before even considering a commercial offer, testing in real conditions is essential.
How do we move from a purely conditional system to an AI-based model capable of managing computer vision, understanding its environment and analyzing all avionics data?
It’s a gigantic transition that involves multiple challenges. The first challenge is cultural and technical: at Airbus, we traditionally operate with a requirement of $10^{-9}$, that is to say the search for absolute certainty. Certification authorities demand full transparency on algorithms and complete visibility across the entire technology chain. We are therefore evolving in the most advanced determinism, while AI, by definition, is based on probabilistic models.
In addition, we are discovering new ways of articulating these systems. For example, it is possible to use AI to generate elements which will then be validated by purely deterministic tests. Likewise, AI agents can entirely rely on deterministic tools. In reality, by integrating human control at the right place in the chain, we are able to achieve levels of certainty much higher than what was thought possible with machine learning four more years ago.
For assisted landing with AI (visual landing), how does this differ from a traditional approach?
In reality, we must consider visual landing AI as an additional sensor, and not as a replacement technology. Far from the mantra of saying it will replace the pilot, this approach complements existing systems, such as ILS or GPS, to provide a layer of redundancy and increase the overall confidence level of the aircraft. Like any sensor, this system has its own levels of certainty, quality and reliability, which must be analyzed and adapted according to operational conditions.
The real challenge for our teams lies first in the transparency of the underlying algorithms, but also and above all in our ability to integrate this technology directly into the device. There is a major technical gap between running an AI in the cloud on the ground and safely embedding it in an avionics system in mid-flight. This research work, carried out for several years, will continue in the long term. This is a complex subject, but it represents a major strategic differentiation area for Airbus.
On a more terrestrial and immediate level, you have just signed a partnership with Mistral AI for your generative AI strategy. Why did you choose Mistral rather than another market player?
Our choice fell on Mistral AI for three fundamental reasons. The first is that it is today one of the best options for deploying sovereign AI in Europe. This does not mean that we abandon American technologies, which we continue to use where it is relevant. Mistral opens the doors to the processing of our highly sensitive data and our critical intellectual property. This will notably make it possible to develop concrete use cases in the defense sector, such as a coding assistant for military applications. Our defense subsidiary was also among the early adopters of their solutions.
The second strong point lies in the size of Mistral, ideal for addressing the very specific needs of Airbus. Training so-called frontier AI models is not our core business: we are consumers. On the other hand, going beyond simple fine-tuning to customize these models with ultra-precise data from Airbus constitutes a real competitive advantage. It’s a two-way partnership: for Mistral, it’s the opportunity to access new industrial data, far from the web standards that have fueled LLMs until now. Ultimately, our ambition is to go beyond the strict framework of LLMs to explore with them the very promising field of physical AI.
A few weeks ago, the Mistral AI teams entrusted us with working on models applied to industrial design, particularly for the design of parts. Is this a technological breakthrough that you plan to concretely integrate at Airbus?
Quite. This is a major area of research that we wish to explore with Mistral, particularly in terms of certification. We want to use AI to automate part of these processes, validate design elements and define, with the authorities, the framework of what is achievable. The real Holy Grail, however, remains the transformation of our upstream design methods. Today, calculating fluid dynamics requires phenomenal power. These simulations are so long and complex that an engineer cannot run billions of them on the fly.
Physical AI is a game changer. In the same way that an LLM assimilates text, these models feed on sensor data and past simulations to learn how to generalize. By streamlining the initial design phases, we will be able to explore many more options. The real value of AI lies not in cost reduction, but in its ability to accelerate the design of new forms of aircraft that are impossible to physically test within traditional time frames. This is an unprecedented competitive advantage for the final product.
We talk a lot about digital sovereignty. If it goes through Mistral and models based in France or Europe, it also depends on the hardware and chips. Do you feel the need to adopt European suppliers?
The choice is particularly limited and we still lack the perspective to make a definitive decision. On the other hand, our ambition is clear: the entire technological stack must, ultimately, become sovereign. For the moment, we are deploying Mistral on our own infrastructures, which certainly rely on American or Taiwanese chips, but we are at the same time considering access to the Trusted Cloud to have sovereign environments capable of ramping up our calculations.
We are realistic: if these systems become widespread, maintaining such computing power internally will become difficult, especially since it is not our core business. It is essential to rely on a European ecosystem and industrial fabric. The ideal would be to go all the way, from energy management to on-board computing. The chip is the final step in this process, and it’s a challenge on a whole new level.
