Industrie - Accélérer la sortie de l’IA des laboratoires pour l’opérationnaliser au service des forces : l’écosystème cortAIX de Thales
Translated with DeepL.com (free version)
Industry - Accelerating the emergence of AI from the laboratory and putting it to work for the armed forces: the Thales cortAIX ecosystem
With around a hundred products already in its catalogue incorporating artificial intelligence (AI), Thales now intends to speed up the process and industrialise embedded AI in critical defence systems. To meet this challenge, and also to make it easier to understand the currently fragmented internal resources, Thales recently launched the cortAIx ecosystem. This is both an informal organisation to enable better sharing and, in some respects, also a managerial organisation via entities concentrating expertise:
- cortAIx lab: a strike force that can be mobilised across the board on certification, qualification, intellectual property, cybersecurity and other issues, located mainly on the Plateau de Saclay (close to the research part of the recently announced Ministerial Agency for Artificial Intelligence in Defence (AMIAD) of the French Ministry of Defence);
- cortAIx sensors: to combine AI, systems engineering and materials science (on radars, sonars, radios, optronics, etc.), and provide the best possible response to the issues of energy efficiency, on-board capability and the integration of software and hardware into sensors right from the product development stage;
- cortAIx factory: a technology factory (in Paris, Rennes, Singapore and Canada) for the acceleration and qualification of solutions in the various decision support use cases, whether in mission planning, UAV/robot piloting, Command & Control (up to C6ISTAR...), etc... ;
A number of partnerships are envisaged or already underway (with start-ups, research centres, manufacturers, etc.), including a very strong desire on the part of Naval Group to work with the cortAIx ecosystem, under arrangements yet to be finalised. All in all, AI is now leaving the research laboratories and getting ready to be used in operational conditions in a number of areas.
AI for maritime patrol radar operators from 2025
From early 2025, AI bricks will be integrated into the Searchmaster radars supplied by Thales and installed on board the French Navy's ATL-2 maritime patrol aircraft. This radar upgrade is being carried out under an 'Other Armament Operations' (AOA) contract, which is different from the contract to upgrade a fleet of 18 ATL-2s to Standard 2. The requirements of the customer (Direction Générale de l'Armement / Marine Nationale) included automatic and intuitive calibration of the radars according to the missions (a tedious process until then), faster detection of targets of interest in the mass of data sent back (to limit the cognitive load of the on-board radar operators, thanks to optimisation algorithms and deep learning) and the fact that these radars are learning (by taking into account the acceptances or refusals of the operators to the proposals made, via reinforcement learning).
The solution developed makes it possible to suggest, in real time, the tracks (air/land/sea) that should be followed, by pre-classifying them. A special symbology (colours/shapes) is used to pre-filter them according to their size, correlation with certain lists (of suspect vessels/ships of interest), possible infringements linked to the absence of self-declaration (AIS type, for example), etc. In 2023, the DGA, the French Navy and Thales conducted a campaign to test a demonstrator over large areas, achieving significant gains in both detection and classification.
These AI bricks will also be standard on the radars aboard the Albatros maritime surveillance aircraft (based on Dassault Aviation Falcon 2000s, replacing the Falcon 50 and Guardian), the first deliveries of which are scheduled for 2027. The naval versions of the HIL (Hélicoptères Interarmées Léger - Light Joint Helicopters) will also be able to work together with the Airmaster axis radars in the nose and flanks of these helicopters.
In 2026, the TALIOS (Targeting Long-Range Identification Optronic System) optronic reconnaissance and targeting pod fitted to the Rafale aircraft will also be 'AI inside'. Until now, the images gathered by scanning a given area using a joystick were analysed in flight by the crew via video feedback, or on the ground once the hard disks had been recovered after landing.
An on-board Thales Neuronal Processor should enable real-time analysis for the detection of a type of target, without going as far as the precise identification of potential targets detected: "it will be 'it's a tank', but we're not at 'it's this type of tank'". According to Thales, the gains achieved are a factor of 100, with the time taken to scan an area to be targeted dropping from 15 minutes at present to a few tens of seconds tomorrow, after an automatic scan enabling the given area to be scanned while maintaining high resolution (a few tens of cm per pixel). The targets will be pre-pointed to the pilot or crew (pilot/copilot), who will confirm the identification and decide for themselves what to do next. The algorithms have been trained using images and data collected both during industrial flights and Air Force and Space Force flights. The other solution is the use of synthetic data generated by Thales via AI, which is combined with data from other sources, while remaining "smart data rather than big data".
According to Thales, its 'sensor' expertise has enabled the integration of an analysis processor (a neural network) that is energy-efficient (a particularly limited resource on board), relatively light and compact, just behind the optronics section. Thales has taken the sovereignty aspect relatively far for this equipment, which is important for the missions carried out (because it enables targeting). The company takes sub-systems from abroad (particularly the United States) and strips them of their software layers, keeping only the hardware and rebuilding their own software layers on top: "An expensive choice, but necessary for certain sovereignty constraints".
This solution will be integrated into the Rafale's F4-3 standard, and will have to be compatible with the current IT infrastructures planned in the fighter squadrons, in particular to train the ground algorithms on the most recently acquired data and to guarantee harmonisation of the algorithms within the fleets. These are complex issues to tackle via clouds, but they should enable the most up-to-date algorithms to be loaded into aircraft for each mission, depending on the preparation phase. As is already the case today with a range of system configurations equipping the Rafale.
AI in your ears soon on board reconnaissance aircraft or on the radio
Other avenues of development are being explored, in particular using algorithms from the civilian sector, such as the denoising of radio conversations on board surveillance or reconnaissance aircraft (ALSR/VADOR type) carrying out long missions (of the order of 10 hours) or for the CONTACT/SYNAPSE range of radios (the export version). The aim is to remove parasitic noise, enabling greater concentration during long sequences interspersed with loss of links or disrupted links, and sound environments saturated with explosions, engine noise, etc.
In the short term, the AI solutions can run on on-board servers (as part of an approach co-developed with Air Force and Space Force users), and in the medium term the processors will be able to be implemented directly in Bluetooth radio headsets. The real added value of the solution will not necessarily be in the algorithm (derived from the civilian sector, to which performance layers have been added, for example to take account of explosions), but in its integration in a critical environment. This requires processors that consume little energy (the real limitation on radio sets), are light and robust.
So the AI needs to be assimilated to specific hardware, via special reinforced silicon wafers (3 mm by 3 mm). Demonstrators of very low-power neural networks have been developed using chips produced by GreenWaves Technologies (in which DGA and Thales are shareholders). And all in compliance with the demanding MIL and aeronautical standards.
Enough to meet the challenges ahead
So there is no shortage of projects for the 600 or so Thales experts (2/3 of whom are in France) working on artificial intelligence (AI) issues. Engineers and researchers, including (according to the company) neither data annotators nor the 100 or so in-house PhD students also working on these subjects. This strike force is responsible for filing an average of 40 patents a year in Europe. This compares with AMIAD, which is aiming to bring together 300 people over time, or other European AI pureplayer companies (in the news in recent weeks) with around 200 to 300 employees, all departments combined..
Beyond that, Thales is seeking to build on what it describes as the company's 'magic square':
- Business knowledge based on experience of the concepts of operations and the environment in which the products developed are used;
- The fact that AI is seen as an 'enabling' technology that is applied to products that are also mastered, and not for its own sake, and where AI and materials science are combined (enabling, for example, the consequences of the use of certain ceramics in sonars on the data collected and subsequently processed to be mastered);
- The search for the most complete control of the entire chain (see the case of the TALIOS pod);
- Cyber security 'by design', drawing on the Group's 5,800 cyber security experts (including the CESTI - centre d'évaluation de la sécurité des technologies de l'information - in Toulouse), and in particular its 'Friendly Hacking' capabilities.
These are just some of the assets that will enable Thales to become a leading player in embedded defence AI in Europe. For Thales, this type of AI is hybrid (symbolic/generative), on a trusted cloud (announced as qualifying by ANSSI at the end of the year), explainable (transparent) and stops before a decision is made (for ethical reasons). To achieve this ambition, there are still challenges to be met, particularly in the development cycles for solutions and in maintaining them at the highest level of the state of the art throughout their lifecycle. There will be no shortage of changes in methods (but also, on a more economic level, in business models), if we are to pass a new milestone after that of digitisation, and fully achieve that of the operationalisation of reliable AI in the service of the armed forces.
