AI and the verification of autonomous systems

As artificial intelligence (AI) - particularly foundation models and generative AI - becomes increasingly integrated into autonomous systems, the robotics community faces fundamental questions about verification, safety, and reliability. While AI offers unprecedented capabilities in perception, planning, and control, it introduces novel challenges: unpredictable failure modes, vulnerability to distribution shift, and opacity that complicates traditional verification approaches. The goal of this workshop is to bring together researchers, practitioners, and industry experts to critically examine the current state of AI-enabled autonomy verification, addressing both successes and persistent gaps. Unlike previous IROS workshops focused narrowly on formal methods or specific AI architectures, this workshop takes a holistic view across the verification landscape. We explore: (1) emerging best practices for verifying large language model (LLM)-based planners and vision-language models in robotics; (2) the interplay between classical model-based verification and learning-based components; (3) whether AI fundamentally constitutes a new class of uncertainty requiring novel frameworks; (4) the implications of this new paradigm in
safety critical systems and (5) identifying scenarios where traditional autonomy
without AI may remain preferable from a verification standpoint. The workshop will feature invited speakers and panel discussions on open problems, case studies of real-world deployment failures and successes, and collaborative sessions to map research priorities. By bridging formal methods, machine learning safety, and robotics deployment experience, we aim to establish a shared understanding of verification challenges and catalyze interdisciplinary solutions.