Robust Design of Trustworthy Neural Network Controllers

Descrizione del progetto

Rendere sicuri i controller delle reti neurali per i sistemi critici

I controller a rete neurale possono ottenere ottimi risultati nelle simulazioni, ma il loro comportamento in sistemi reali, critici per la sicurezza, attualmente non può essere garantito in modo affidabile né verificato formalmente. Questa mancanza di certificazioni in termini di sicurezza e robustezza ne limita l’utilizzo in settori ad alto rischio come quello aerospaziale. Il progetto RDTNNC, finanziato dal programma di azioni Marie Skłodowska-Curie, si propone di rendere i controller basati su reti neurali sicuri, affidabili e certificabili per l’uso in sistemi ad alto rischio. Per farlo, svilupperà un quadro unificato che combina l’apprendimento automatico con la teoria del controllo per progettare, verificare e interpretare i controller a rete neurale con garanzie formali di stabilità e sicurezza. In tal modo, contribuirà a superare la mancanza di affidabilità dimostrabile, una limitazione fondamentale dell’IA in settori come l’industria aerospaziale, la sanità, la robotica e le infrastrutture critiche.

Obiettivo

The Robust Design of Trustworthy Neural Network Controllers (RDTNNC) project aims to establish a unified framework for designing, verifying and interpreting neural network (NN) controllers with formal guarantees, ensuring the robust stability and safety of neural feedback loops (NFLs). It combines model-based and data-driven approaches to address complexity, uncertainty and data scarcity in artificial intelligence (AI)-based control. The project focuses on one of the main barriers to applying AI in high-risk domains: the lack of certified trustworthiness. Although NN controllers demonstrate strong performance in simulations and prototypes, their use in fields such as aerospace, healthcare, robotics and critical infrastructure remains constrained. Without methods that provide the level of reliability and transparency required by regulators, adoption in these areas remains limited. RDTNNC aims to close this gap between empirical capability and verifiable trust. To achieve this, the project advances scalable verification methods based on integral quadratic constraints and sum-of-squares programming, integrating them with Lyapunov-constrained training to embed robustness and stability in controller design. Data-efficient strategies using meta-learning and few-shot adaptation aim to enable safe synthesis even when models are incomplete or data are scarce. By delivering controllers that combine high performance with verifiable reliability, RDTNNC aspires to contribute to the wider adoption of AI in safety-critical environments. The outcomes are expected to support the objectives of the Horizon Europe Work Programme by promoting trustworthy and human-centric digital technologies, by aligning with the requirements of the EU AI Act for high-risk applications, and by strengthening Europe's role in the responsible development of advanced AI-based control.