VAIPOSA advances verifiable AI for resilient PNT

VAIPOSA (Verifiable AI Positioning Applications) has recently been successfully completed, marking an important step in exploring the use of Artificial Intelligence techniques to improve the robustness and reliability of Positioning, Navigation, and Timing (PNT) solutions. Funded by the European Space Agency (ESA) through its NAVISP programme, VAIPOSA has been led by FBK (Fondazione Bruno Kessler), with the participation of GMV.  

The project investigated how AI-based navigation methods can be applied while maintaining the reliability, traceability and safety awareness required for safety-critical applications. VAIPOSA demonstrated that combining GNSS (Global Navigation Satellite System) and SLAM (Simultaneous Localization and Mapping) through a safety-supervised framework can improve robustness compared with standalone navigation solutions, particularly in challenging environments. GMV contributed to the navigation architecture by integrating a virtual, high-fidelity GNSS PPP receiver into the open-source CARLA autonomous driving simulator, enabling GNSS-based positioning to be exploited with AI-based SLAM techniques. 

The experimentation campaign carried out as part of the project involved autonomous driving simulations covering different types of urban environments, as well as varying weather and visibility conditions. A central outcome was the development and assessment of the so‑called Safety Cage concept, which monitors the behaviour of the navigation components, assigns reliability weights and supports mitigation actions when anomalous behaviour is detected. 

Important challenges remain, including the integration of additional sensors, such as IMUs (Inertial Measurement Units), and the need to further investigate integrity in multi-sensor navigation systems. When GNSS, SLAM and other sources of positioning information are fused together, it is not sufficient to assess each component independently: it is necessary to understand how their individual uncertainty, confidence and integrity information should be combined and propagated at system level. Defining such a unified multi-sensor