Surrogate-Assisted Search-Based Testing of Robot Controllers Based on Enzymatic Numerical P Systems

Surrogate-assisted search-based testing (SA-SBT) reduces simulation cost by replacing expensive runs with lightweight mathematical or data-driven approximations to efficiently guide metaheuristic search toward safety-critical system failures. While existing SA-SBT techniques predominantly employ black-box surrogates evaluated on short, static scenarios, they rarely exploit internal controller specifications or assess transferability to reactive closed-loop environments. In this paper, we aim to investigate how SA-SBT can be applied to a robot controller based on Enzymatic Numerical P systems (ENPS). We build and evaluate two surrogate strategies: a black-box classifier using road geometry and ENPS model identifier and a gray-box surrogate reconstructed directly from the controller internal rules. The classifier substantially reduces the number of required simulations while retaining nearly all true failures on independently generated test roads. In addition, we present and evaluate the gray-box surrogate, examining whether continuous fitness prediction, the mechanism underlying prior SA-SBT surrogates from the literature, transfers to a controller governed by a long, reactive control loop.