Testing Cyber-Physical Systems (CPS) is crucial, as they play a central role in modern society and are deployed in several domains such as autonomous driving, robotics, and smart infrastructures. These systems integrate software controllers with physical processes, leading to highly complex input spaces and nonlinear behaviors. As a result, even small variations in operational conditions can cause large deviations in system performance, often degrading Quality of Service (QoS) or even leading to safety violations. Identifying such \textit{boundary scenarios}, inputs where minimal perturbations lead to large behavioral changes, is fundamental to ensure high system dependability prior to deploying a software version in operation.

However, generating boundary scenarios for CPSs poses several challenges: (i) the simulation-based execution of CPS test cases is computationally expensive; (ii) the input space is multi-dimensional and includes time-dependent signals; (iii) QoS requirements are difficult to formalize due to the complex outputs; and (iv) many automatically generated test inputs may be physically invalid. While prior work has addressed these challenges individually, no comprehensive solution exists that combines the test input minimization, behavioral boundary detection, and test input validity.

In this experience paper, we propose LiftJanus, the first search-based test generator for CPSs that integrates three state-of-the-art techniques into a unified framework. First, a test input minimization component that by means of delta-debugging reduces the test inputs while preserving a set of pre-defined properties. Second, a boundary detection algorithm, inspired by DeepJanus, generates input pairs that differ minimally but induce maximally different QoS outcomes, thereby exposing sensitive operating boundaries. Third, an automated repair component adjusts CPS configurations, ensuring that the produced test cases are physically feasible.

We empirically evaluated LiftJanus in two real-world elevator systems provided by our industrial partner, \Orona, one of the largest elevator manufacturers worldwide. The evaluation combined quantitative and qualitative analyses. Quantitative analysis showed that LiftJanus generated boundary inputs that were on average twice as effective as baseline techniques in degrading system QoS. The repair mechanism successfully improved configuration parameters in 76.25% of boundary cases. With respect to the qualitative assessment, open-ended interviews with domain experts confirmed that LiftJanus not only identifies scenarios previously overlooked by manual testing but also provides actionable insights for improving system configurations and dispatching algorithms. Experts stated that generating minimal yet impactful test cases is particularly valuable, as it reduces analysis overhead. Moreover, we grasped a set of lessons that can be beneficial to other practitioners in different domains.

In summary, this paper makes three key contributions: (1) We introduce LiftJanus, a novel comprehensive CPS testing framework. (2) We demonstrate, through an industrial case study, that LiftJanus scales to complex, real-world systems and produces significantly more effective and valid boundary test inputs than baselines. (3) We report practical evidence from domain experts, highlighting both the benefits of boundary-based testing and the challenges of transferring research prototypes into industrial practice.