Nathan Lichtlé: Developing autonomous driving algorithms to optimize traffic flow

• Thesis Title: Stabilization and Control of Dynamic Systems: From Classical Methods to Reinforcement Learning (https://theses.fr/2024ENPC0030 / https://theses.hal.science/tel-05004078)
• Awarding institution: École nationale des ponts et chaussées, in joint supervision with the University of California, Berkeley
• Thesis advisors: Amaury Hayat (CERMICS) and Alexandre Bayen (University of California, Berkeley, EECS, BAIR)
• Current position: AI startup

What motivated you to write a thesis on socio-economic issues?

During my studies, I completed several research internships that inspired me to pursue a Ph.D. to delve deeper into topics that have long interested me, particularly in artificial intelligence and control systems. Pursuing a Ph.D. allowed me to work at the cutting edge of research in these fields while retaining the freedom to choose my own projects—something that’s difficult to achieve in industry. However, I was keen to work on practical applications. Since I already knew my advisors from my last internship, I knew I would be able to collaborate with the automotive sector and conduct experiments on vehicles under real-world conditions—which was exactly what I was looking for.

You received the Caseau Prize for the theme “Decarbonizing the Economy.” Can you tell us about your dissertation topic and how it relates to this theme?

I worked on many different projects during my PhD, but my main focus was on improving traffic flow by intelligently controlling autonomous vehicles operating alongside other drivers. I developed reinforcement learning methods, at the intersection of artificial intelligence and control theory, as well as traffic simulators. This allowed me to design driving algorithms capable of learning, from millions of simulations, to drive more smoothly and alleviate traffic congestion. I have also optimized these approaches to minimize fuel consumption and CO₂ emissions on a large scale. My thesis therefore demonstrates that a small proportion of vehicles driving intelligently is sufficient to stabilize traffic and generate significant fuel savings for all users, with a direct impact on the carbon footprint. Finally, I validated these results on a large scale by deploying my algorithms on approximately 100 autonomous vehicles on the highway, in collaboration with the CIRCLES consortium and industrial partners, including Nissan, Toyota, and GM.

What were the main challenges you faced during your research?

My main challenge during this project was to make the driving algorithms robust enough for deployment in real-world conditions. The problem is inherently complex: it involves developing algorithms capable of controlling a very small number of vehicles in a way that influences the entire traffic flow—a dynamic system where hundreds, even thousands, of vehicles interact with varied behaviors. The major challenge was to go beyond the scope of simulation and design methods directly transferable to a highway involving 100 autonomous vehicles, in heavy rush-hour traffic, with strict safety constraints. What makes this difficult is that reality is far more complex than simulations: vehicle sensors are not perfect, drivers may behave unpredictably, and the algorithms must be robust against these disruptions. Over the course of three years, I first developed numerous prototypes to validate the approaches in various scenarios, then fast and realistic traffic simulators based on real-world data to better capture the phenomena observed on highways. Finally, the most difficult step was the transition to reality: integrating these methods into commercial vehicles while ensuring their reliability and safety.

In what ways has your work benefited the socio-economic world today, or will it benefit it in the future?

There is already a tangible impact: the project has shown that by controlling a very small proportion of vehicles (around 1 to 3%), it is possible to smooth out traffic flow, reduce congestion, and lower fuel consumption for all road users. On a highway-wide scale, this can amount to several thousand liters of fuel saved each day. From a socio-economic perspective, the approach is appealing because it does not require changes to road infrastructure: it can be integrated into existing driver-assistance systems found in the vast majority of modern vehicles via simple software updates, with potentially very low deployment costs. Drivers could then activate this type of system on the highway, just as they already do with adaptive cruise control. By comparison, traditional traffic management solutions often rely on infrastructure that is costly and time-consuming to implement, such as variable speed limit signs or traffic lights at highway entrances. On the industry side, collaborations with Nissan, Toyota, and GM have validated the feasibility of these approaches, and some of their research labs are already experimenting with the algorithms developed during the project.

What advice would you give to young people who want to focus their research in mathematics on practical applications?

I would recommend looking early on for a research framework that bridges theory and practical applications—such as a CIFRE thesis or a research team collaborating with industry partners—and choosing your advisors accordingly, since these collaborations often determine access to hands-on experiments. In my case, without industry partners, I might have been able to conduct experiments on one or two vehicles, but never on a scale of 100. Doing internships in industry during your thesis is also a great way to gain this experience and work on interesting applications. I also think it’s important to start with a real-world problem rather than a method, which allows you to move away from pure theory by first quickly testing methods and ideas on simple systems to assess their viability. This is all the more important because once you start working on concrete applications, everything becomes much more complex and less “clean” than in theory (noisy data, fragile assumptions, unforeseen constraints…), hence the value of confronting these challenges as early as possible.

How do you see the role of mathematics in addressing current societal challenges?

I see mathematics as an essential tool for understanding and influencing complex systems. Many of today’s societal challenges—such as transportation, healthcare, and energy—are driven by collective dynamics that can be difficult to grasp when human intuition reaches its limits. Mathematics allows us to model these systems, and thus to understand them, anticipate their effects, and test measures before implementing them. For example, in my field, it helps us understand the counterintuitive mechanisms of traffic instability—specifically why waves of acceleration and deceleration often form on highways without any accidents or obstacles to explain them—and design robust control strategies to address them. For me, mathematics and artificial intelligence are complementary: AI provides flexibility and enables navigation of highly complex systems, while mathematics provides the framework, the understanding of the system, and the necessary safeguards to move toward solutions that can be safely deployed. Beyond producing models, they also allow us to quantify uncertainties and understand limitations, which is often essential when important decisions depend on them. More broadly, I believe they help develop a way of reasoning and a critical mindset that are very useful beyond the scientific realm, for tackling complex problems in many fields.

What are your plans for the future?

Today, I’m working on several artificial intelligence projects at a startup, particularly in the medical field. The two things I enjoyed most during my PhD were building concrete, high-performance systems—such as simulators and machine learning algorithms—and training artificial intelligence systems to watch them learn and evolve, particularly in the simulations I created for them, but also in real-world conditions, such as when they simultaneously controlled a hundred vehicles on a highway. I continue, in line with my PhD work, to design and build systems for very practical applications at the cutting edge of artificial intelligence research.