Meet Gauthier Picard, a Senior Research Scientist at ONERA, the French Aerospace Lab. Based in Toulouse, Gauthier is part of the Department of Information Processing and Systems, where he focuses on decision-making, optimization, and artificial intelligence.
As Co-Head of the Artificial Intelligence Lab at ONERA, his expertise spans from classical AI and machine learning to multi-agent systems and optimal coordination. His work is applied to a variety of aerospace-related projects, including collective robotics, satellite operations, and aero-terrestrial robots. In the Domino-E project, Gauthier tackles complex decision problems for managing satellite constellations.
What role do you play in the Domino-E project, and how are you using AI to improve satellite image acquisition?
“Imagine you want to take pictures of different parts of the Earth using satellites. The challenge is figuring out which satellite should take each picture and when it should do it. There are many satellites moving around the Earth, and they can point their cameras in different directions.
The goal is to make sure that the satellites work together in the best way possible to take all the pictures as quickly as they can. So, you have to decide which satellite takes each picture and at what time, so everything gets done as fast and efficiently as possible.”
Can you explain the purpose of the Coverage Service Domino in Domino-E? How does it change the way we approach satellite imagery?
“Imagine you want to acquire images of a large area on Earth. Then you have two options: rely on a single mission or query multiple missions in parallel. Relying on just one mission may take a long time, while querying multiple missions can speed up the process but comes at a higher cost due to data redundancy and wasted images.
To address this, instead of manually requesting data from several missions, an automated approach can be used. This involves splitting the large area into sub-regions, assigning each sub-region to a different mission, and distributing the workload to get images faster.”
What’s complicated about breaking down a large area into smaller regions?
“Satellites are like machines in a factory—they have limited active time and can only operate at specific moments. The goal is to schedule these satellites to capture images as efficiently as possible, much like assigning tasks to machines to produce a final product.
In this case, the large area images you want are the product. You break the area into smaller parts and assign each part to the satellite that can handle it best. The challenge is to coordinate the satellites to minimize time and maximize image quality by considering factors like weather conditions and angles.”
So this requires a lot of decisions to be made. What is your main objective when it comes to solving these complex decision-making problems?
“Our job here at ONERA is to turn complex problems into decision frameworks, like optimization or constraint programming, and then find the best algorithm to solve them. For example, in the Domino-E project, we use a several constraint optimization approaches.
Choosing the right algorithm is a balance. High-quality solutions often take more time, so we look for algorithms that can provide good results quickly, even if they’re not perfect.”
What novelty is required from your algorithms?
“It’s the algorithm itself. The algorithm’s search space, which includes all the combinations that need to be analyzed, grows exponentially with finer subdivisions. The finer the splitting, the larger the search space becomes, making it more difficult to find the optimal solution.”
Can you give us a practical example of how your algorithms work, perhaps by relating it to a real-world problem?
“A useful analogy is to treat the satellite’s task of capturing images as a version of the ‘traveling salesman problem.’ In this classic problem, a salesman must visit several cities while minimizing travel time.
For satellites, the goal is similar: they need to capture images at different points on Earth while minimizing the time spent moving between locations. By mapping the satellite’s task to this well-known problem, we can use proven algorithms to find efficient solutions, reducing the time the satellite spends not taking pictures, while changing his location.”
How do your algorithms handle situations where a satellite might not be in the perfect position to capture an image?
“When a satellite isn’t in the ideal position, it must capture images from a different angle, affecting image quality. The two main factors influencing image quality are the angle of acquisition and weather conditions.”
So, the algorithm decides …
“… what image a satellite takes and when. Because any point on Earth can potentially be imaged by a satellite, but it depends on the satellite’s position and timing. Even agile satellites, which can maneuver to target different areas, require specific overflights to capture images.”
What are the typical use cases you think of?
“Most of the time, you need to acquire images of large areas, often for long-term observations. For example, if you’re monitoring the polar regions to track ice surface reduction, you need to capture the area repeatedly over an extended period to analyze changes. In contrast, during an emergency like an earthquake, the focus shifts to quick access to data about the disaster’s impact. Such disasters typically cover smaller areas, often only hundreds of square kilometers, which can be managed relatively quickly compared to vast regions like an entire country.
Countries like France or Germany use these services to monitor various aspects of their environments, such as forest health, urbanization, and land use. They analyze these large-scale images to gather statistics and make informed decisions about urban planning and environmental policies. Monitoring can also include endangered zones, ocean temperatures, and sea levels to track long-term changes in climate.
For example, during summer, these countries might focus on the impact of forest fires, observing large areas like the south of France to determine how many square kilometers were affected. This helps them gain a comprehensive understanding of the environmental impact over vast regions.”
How do you monitor and track rapidly evolving phenomena on Earth using satellite technology?
“For example, when monitoring phenomena on Earth that require frequent observation—like biodiversity changes, wildfires, or urbanization growth—you need to consistently pass over these areas. If a specific event starts to deviate, like a wildfire spreading in a new direction, you want to follow its progression closely.
This requires re-tasking your satellites to focus on the changing situation. Our work involves broad monitoring to maintain a constant flow of information, but also zooming in on specific events when they emerge. This approach ensures that we have the most accurate model and understanding of the dynamics of these phenomena.
When dealing with rapidly evolving events, speed is critical. It’s essential to react quickly and acquire data at the same pace as the phenomenon itself evolves. If the event covers a large area, you need the capability to gather information on a large scale in a highly responsive manner.”
What’s your long-term vision for the Coverage Service Domino? How do you see it evolving to meet the future needs of satellite missions?
“The goal is to reduce the time needed to capture a large area. For example, if one mission takes a month to cover the area, adding another mission with the same capabilities could cut that time in half, assuming weather conditions are identical.
However, weather remains a major uncertainty that affects all satellites equally.
Currently, we focus on optical sensors, but other sensor types, like radar or infrared, could be used when weather or light conditions make optical imaging difficult. Our framework can adapt to these different scenarios, allowing for things like radar-based imaging when clouds are present.
In the future, we may expand to handle more advanced use cases, such as stereoscopic imaging for 3D data. While our current focus is on optical missions, our algorithms are designed to support these evolving needs.”
How much of your work is theoretical versus applied research?
“We’re conducting both fundamental research and applied development on algorithms for optimization, solving decision problems, and learning problems. We model these problems, create algorithms, and evaluate their performance using real-world data.”
Thank you!
