Quantifying AI failures for integration into physical AI reliability models

Physical Artificial Intelligence (Physical AI) brings learning-enabled functions into systems that interact directly with the physical world. Examples range from autonomous robots and mobile platforms to industrial machinery and road vehicles. Reliability of the physical part of such systems is commonly addressed using established methods based on failure rates, reliability models, diagnostics, redundancy, and common-cause analysis. The behavior of AI components is treated differently. AI-related failures are usually evaluated through testing, scenario coverage, verification, operational monitoring, and safety argumentation, which makes their direct integration into classical system reliability models difficult. This paper proposes a quantitative representation of AI behavioral failures that can be used together with conventional reliability engineering. An AI failure is defined as an observed violation of an allocated automation task under specified operating conditions. The occurrence of these violations is represented by a counting process. Under a non-homogeneous Poisson model, the corresponding failure frequency is shown to be equal to the failure rate of an equivalent AI reliability block. This makes it possible to express AI behavioral performance in failures per hour and, where useful, in equivalent PFH or FIT units. AI components can consequently be included in reliability block diagrams together with hardware, mechanical, and conventional control-system components. AI updates are represented as behavioral renewal epochs, while a βAI factor is introduced to account for the continuous monitoring and inclusion of edge cases appearing in the long tail of distribution with AI verification. Analytical derivation and representative examples illustrate the applicability and limitations of the proposed approach.