087 - Verifiable AI/ML techniques for PNT applications

087 - Verifiable AI/ML techniques for PNT applications


Artificial intelligence (AI) is increasingly utilized in the PNT field, enhancing the performance and reliability of Global Navigation Satellite Systems (GNSS), from RF interference detection and mitigation to sensor fusion. The advancement and ever-increasing size of neural networks increase the complexity of applications supported by AI, and with the increasing complexity, verifiability decreases. Unreliable or biased AI systems can have serious consequences, risking damage both to the environment and to human lives, especially in domains like autonomous vehicles, UAVs, and marine applications. 

To overcome concerns about data biases, insufficient data, and lack of transparency, effective techniques such as data augmentation, transfer learning, and Explainable AI (XAI) methods can be employed. However, addressing issues like inaccurate specifications and algorithmic biases is even more crucial to avoid incorrect and undesirable outcomes. Therefore, it is essential to develop methods that can verify the behaviour of AI systems. Verified AI, as defined by the Association for Computing Machinery (ACM), seeks to design AI systems that provide robust assurance, ideally with provable correctness based on mathematically specified requirements. This approach is crucial to ensure reliable and trustworthy operation of AI systems in PNT applications. 

The objective of the activity is to study the viability of integrating verifiable artificial intelligence and machine learning techniques to enhance PNT applications. 

The tasks to be performed shall include: 

  • Review of the state-of-the-art verifiable AI techniques and design processes for ML systems in PNT domain 
  • Develop verification algorithms based on formal specification and formal verification methods to enhance trust and verifiability in AI systems operating in the PNT field, where the human-AI interaction at design level is instrumental in ensuring reliable and transparent performance for the human agent. 
  • Develop a simulator to test few specific use cases in complex environments (e.g., sensor fusion) 
  • Design a demonstrator to evaluate the performance of the developed verified AI algorithms for PNT.


The main outputs of the activity will consist of: 

  • Design and development of verifiable AI/ML architectures and algorithms applicable in the PNT domain for autonomous systems (e.g., maritime, road, railway, or aviation) 
  • Simulator and demonstrator to assess the verified AI algorithms in challenging environments. 
  • Recommendations for best practices for testing AI for PNT applications