Thales engineers are using an uncrewed air vehicle to test AI in navigation.
Thales is developing hybrid artificial intelligence approaches for commercial air traffic management and large drone control, thus taking steps toward eventual use on commercial aircraft.
Thales aims to harness the power of artificial intelligence (AI) in future commercial cockpits, with possible advancement toward autonomous flight. Engineers have found ways to circumvent AI’s challenges for aviation: explainability, the system’s ability to explain its decision or advice, and trustworthiness, or ensuring AI does not make the wrong decision. Whether crew members or air traffic controllers can maintain the current level of engagement remains to be seen. While the pilot or controller is making the final decision, they do not devise the solutions themselves.
- Drone flight-testing advances AI-based navigation safety
- AI may help maintain trajectory separation in air traffic management
Thales’ progress on AI comes as the European Union Aviation Safety Agency (EASA) has issued rules intending to allow certification in 2025 of a Level 1 AI system, which provides assistance only while a human operator remains in command. A number of aerospace players are working on this. Daedalean and Leonardo have flight-tested AI tools to support helicopter pilots with traffic detection, navigation and landing. Daedalean’s research in 2020-21 contributed to EASA’s first guidance for Level 1 AI applications in aviation.
Thales recently launched its Autonomy Project, Baptiste Lefevre, Thales Avionics’ technical lead for autonomy, said at the AI Applications in Aeronautics, Defense and Space conference organized by the Air and Space Academy in Paris in mid-November. “We want to use the amazing optimization power of AI while keeping decades of experience in safety management and safety analysis,” he emphasized. Thales has thus set a prerequisite to achieve the project’s goal of full autonomy: a higher safety level than current aircraft.
Under a stepped approach, Thales engineers have started to mature solutions on a long-range UAS100 surveillance drone. They deem urban air mobility the logical next target due to the high relative cost of pilots in operating an aircraft carrying a small number of passengers. After 2040, Thales intends to pursue full autonomy of a commercial aircraft.
“That means maybe 2050 or maybe never, but it is still our horizon,” Lefevre said. “Our deliberate choice for a long-term objective pushes us to find new concepts, new systems and new technologies that will drive future developments.”
On the UAS100, Thales devised an AI system for a car-tracking mission. The drone’s trajectory depends on a moving target and cannot be planned. “We decided to use a neural network to generate a trajectory, using reinforcement learning,” Lefevre said. “We are calling the neural network every second, giving current positions as inputs. As the output, the system gives a direction and a radius of curvature.”
The integrity of an AI system is below 99%, far from the required safety level on such an aircraft—the probability of a catastrophic failure should be one in 10 million or below, Lefevre explained.
To bridge that gap, Thales has adopted a hardware architecture relying on two onboard computers. The first uses high-performance processing units able to run a neural network, but it is limited to a relatively low reliability level. The second meets safety-critical criteria with a high design assurance level but has moderate performance, Lefevre said.
From the ground, the operator sends just a mission objective. The high-performance computer receives it and generates a trajectory using AI. It then shares the result with the high-design-assurance computer, which checks the trajectory. “The architecture unlocks free-route capability for the drone,” Lefevre said, referring to the ability to choose a flightpath on its own, without an onboard program or ground control.
Engineers have been very careful about the type of trajectory the high-assurance computer accepts, he said. If it does not match with the drone’s performance or the obstacle database, the trajectory is rejected.
The first computer also generates contingency trajectories. In case of a failure, the second computer checks all the database’s “what if” scenarios. “In the certified computer, at all times, we have this trajectory database for safe termination of the flight,” Lefevre said. Flight tests in Canada included car tracking as well as power line detection and avoidance.
For air traffic management (ATM), Thales is investigating the combination of a neural network with a certified conflict-detection system to help prevent aircraft from flying too close to one another in a given airspace.
The emergence of such a hybrid AI system could reduce pressure on air traffic controllers, who would decide whether to use a proposed resolution advisory. It also could improve avoidance trajectories, for instance, by making them shorter. Thales says such a system might enter service around 2030, after further development and certification work.
Under the conventional procedure, when a trajectory conflict or potential loss of separation is identified, the air traffic controller steps in and identifies a new trajectory. They provide the crew with an instruction, or clearance, to alter their flightpath and avoid the conflict. Simultaneously, controllers make sure they avoid conflicts with other aircraft in the same airspace.
Regulations call for maintaining separations of 5 nm horizontally and 1,000 ft. vertically, although some variations may apply. “When traffic is very dense, this can be quite stressful,” Thales Airspace Mobility Solutions Chief Data Scientist Andrei Purica said at the conference.
Thales already offers an automated resolver. Based on classical research algorithms, it works in the background, looking for an alternative trajectory, Purica said. The resolver shares the result with the controller, who can decide whether or not to use it.
The current model has computational power limits, however, and cannot analyze every single solution in the background. Thus, it may miss some, Purica pointed out.
Thales is exploring AI to address this—specifically, reinforced learning. The approach gives more exhaustive and faster results, and includes cost functions, such as the benefits and drawbacks of each solution. It can take into account criteria including CO2 emissions and contrail reduction, Purica noted.
Safety concerns rank first among the limitations of AI for ATM. The system cannot guarantee a solution, and a solution may be difficult to explain. “In accuracy, 99% is not enough,” Purica said.
Hence Thales’ hybrid approach, combining AI and symbolic calculations. “If a conflict is detected, and we have a number of options to resolve it, we ask the neural network what it thinks of a certain option with respect to a certain cost function,” Purica said. “We look at rewards and penalties. In a flightpath’s deviation, the neural network may tell us one heading is better than the other, or it may suggest a certain range of angles is best.” The classical tool then focuses the search and gives the controller a validated solution.
To train the AI model, Thales relied on 20-30 min. ATM scenarios and ensured it was learning from failures, Purica said. The company deemed the experiment successful. “We ran the system for several days,” he said. “After 15 million clearances, the system began to understand and started to maximize the cost function we provided. After 20 million clearances, it not only solved 99% of the problems, it reduced the average length of the trajectories.”
Thales tested the AI part of the system with controllers, who found it acceptable—sometimes comparable to a novice controller, sometimes more proficient.
Related Content
