As 2025 draws to a close, it is the occasion to look back on the progress of the RobustifAI project, launched in June 2025 with a clear ambition to rethink the robustness and reliability of AI systems. Since then, our teams have achieved major breakthroughs:

• 𝗠𝗼𝗱𝗲𝗹𝗶𝗻𝗴 𝗛𝘂𝗺𝗮𝗻-𝗖𝗲𝗻𝘁𝗿𝗶𝗰 𝗖𝘆𝗯𝗲𝗿-𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗦𝘆𝘀𝘁𝗲𝗺𝘀: we collected concrete use-case requirements to identify the most suitable specification and modelling languages for expressing the properties of human CPS and accurately modelling their environments
• 𝗥𝗼𝗯𝘂𝘀𝘁𝗻𝗲𝘀𝘀 & 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗼𝗳 𝘃𝗶𝘀𝗶𝗼𝗻-𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗺𝗼𝗱𝗲𝗹𝘀: we extended randomized smoothing to VLMs, using an oracle to classify generative outputs (e.g., harmful vs. harmless responses). Even with imperfect oracle accuracy, it guarantees robust certification against jailbreak attacks on state-of-the-art VLM.
• 𝗩𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 & 𝘃𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗚𝗲𝗻 𝗔𝗜: new techniques have been developing to ensure the safe and reliable use of AI-generated code by automatically associating code with correctness invariants that can be formally checked using existing verification tools.
• 𝗦𝗰𝗲𝗻𝗮𝗿𝗶𝗼𝘀 𝗳𝗼𝗿 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗺𝗼𝗯𝗶𝗹𝗶𝘁𝘆: scenario-based generative AI agent now learn from vehicle and accident databases, align with standards and regulations (ISO, UNECE, Euro NCAP), and automatically generate human-readable, industry-standard dynamic driving scenarios (XML/DSL).
• 𝗢𝘂𝗿 𝘁𝗲𝗮𝗺 𝗵𝗮𝘀 𝗯𝗲𝗲𝗻 𝗮𝗱𝘃𝗮𝗻𝗰𝗶𝗻𝗴 𝗩𝗟𝗠𝘀 to generate scenario labelling, metadata, and novel image, video driving scenarios. We have also been developing AI Agents for requirements management, supporting consistency checks and conflict detection across complex requirement databases.

A huge thank you to all the project partners and contributors who made these advances possible. The adventure is just beginning!

Caption for the video below: 3D Gaussian Splatting Model Reconstruction (copyrights: Siemens Industry Software NV):