In an AI-enabled CPS, a DNN serves as controller of the physical system. Despite extensive training, a DNN may still produce wrong decisions, causing safety risks. DNN repair is a search-based approach that can be potentially used to perform targeted improvement of the DNN controller. However, existing DNN repair methods rely on ground-truth labels for each inference, which are unavailable in AI-enabled CPSs. Instead, system-level specifications can be used to detect wrong behaviours. Based on this intuition, we propose a novel search-based repair approach guided by system-level specifications. It takes as input a specification, a test suite with passing and failing tests (i.e., satisfying/violating the specification), and some faulty DNN weights; given these inputs, it searches for alternative weight values to improve the performance of the DNN controller. We also introduce a heuristic that accelerates the search. Experiments on real-world AI-enabled CPSs show that the approach effectively repairs their controllers. This abstract summarises: “D. Lyu, Z. Zhang, P. Arcaini, F. Ishikawa, T. Laurent, J. Zhao. Search-Based Repair of DNN Controllers of AI-Enabled Cyber-Physical Systems Guided by System-Level Specifications. In GECCO 2024.”