This paper presents the experimental validation of an automated synthesis method for Augmented Reality (AR) and Virtual Reality (VR) architectures. Unlike traditional analysis methods that evaluate fixed designs, this approach utilizes a Genetic Algorithm (GA) within a Simulink environment to actively generate architectural configurations that maximize resilience. An experimental setup, a mapping of 20 distinct mitigation strategies to design variables, and the results of a simulation under severe operational conditions are described. The experiment demonstrates that the synthesized architecture achieves a significant improvement in overall resilience, with critical gains in availability and recovery time, proving the practical utility of automated design optimization for immersive systems.
As Augmented Reality (AR) and Virtual Reality (VR) scale beyond niche applications into critical
infrastructure for healthcare, industrial manufacturing, and education, their reliability becomes
paramount [
In this context, resilience as the capacity to maintain acceptable functionality despite external
stress or internal failure, is not merely a feature but a fundamental requirement. However, current
engineering practices largely rely on passive analysis, evaluating how pre-determined, fixed
architectures behave under fault conditions [
Building upon our previous degradation analysis framework [
To construct the 20-dimensional design space for the synthesis engine, let us aggregate diverse mitigation strategies from recent literature, mapping them to controllable simulation variables across four architectural layers.
At the hardware layer, the focus was on managing thermal and energy constraints to prevent
physical throttling. The cloud/edge offloading techniques proposed by Sun et al. [
For the software subsystem, mechanisms were selected that enhance recovery speed and
adaptability. The synthesis model utilizes checkpoint-restart frequencies based on the work of
Foerster et al. [
The data and communication layer can be parameterised to optimise throughput and integrity.
Multipath scheduling strategies from Zhao et al. [
Finally, user-centric mitigations can be implemented to directly address safety and physiological
comfort. To combat cybersickness caused by latency, motion prediction and frame interpolation
techniques [
To validate the method, we constructed a high-fidelity simulation environment using MATLAB/Simulink. The experimental setup was designed to replicate a "mission time" of 5000 time units under dynamic stress.
The core of the experiment is a Simulink model (Fig. 1) that simulates the VR system's behavior. The model accepts five categories of dynamic inputs that represent real-world disruptors: network conditions (bandwidth fluctuation, jitter), environmental variables (lighting changes, thermal ambient conditions), user behaviour (erratic motion, lack of attention), system load (computational spikes), and hardware constraints (battery droop, thermal throttling). These inputs feed into a Failure Simulation Block, which triggers probabilistic degradation events based on defined hazard rates.
The "control knobs" for the experiment are represented by a parameterised vector of mitigation
approaches (Mvector). This vector consists of 20 discrete values, each corresponding to a specific
technical countermeasure derived from literature [
The MATLAB Global Optimization Toolbox was utilised to drive the simulation. The synthesis process is formulated as a single-objective optimization problem where the Genetic Algorithm seeks to find the vector mopt that maximizes the aggregated resilience score Rsys.
The evaluate_analytical function. It takes a candidate Mvector, applies it to the Simulink model, computes the seven resulting resilience metrics, and returns a weighted global score.
The GA iteratively refines the population of vectors, "breeding" superior designs by combining mitigation strategies that successfully maintain high resilience scores [28].
To validate the proposed synthesis method, we constructed a comprehensive experimental
framework utilising a MATLAB/Simulink environment, selected specifically for its capacity to
model both continuous-time dynamics and discrete-event failures. The experiment was defined by
a specific "mission time" duration of 5000 time units, during which the system was subjected to a
dynamic operational profile that fluctuated between idle, normal, and peak loading states. The
primary objective of this setup was to solve the inverse design problem: identifying the
architectural configuration that maximizes the aggregated resilience score under specific
constraints [
To rigorously test the efficacy of the synthesis method, we defined a "severe operational scenario" rather than a standard use case, introducing high-frequency disturbances across five distinct categories. The simulation introduced network instability characterized by high jitter and bandwidth throttling to mimic poor edge conditions, alongside environmental stresses such as rapid changes in ambient lighting and temperature These external factors were compounded by erratic user behaviour, including intense motion and high interaction rates, as well as internal system load spikes that threatened computational saturation. Finally, the scenario accounted for hardware constraints by simulating battery degradation and thermal throttling events. This harsh baseline was intentional, ensuring that a system without active resilience mechanisms would fail significantly, thereby highlighting the specific gains provided by the synthesized architecture.
The core of the experiment relied on a closed-loop integration between the Genetic Algorithm (GA) and the Simulink model. The process begins with the GA generating a candidate 20dimensional mitigation vector which represents specific engineering decisions such as redundancy levels or throttling thresholds. These abstract values are then mapped to concrete Simulink parameters that govern the system's physical behaviour during execution. As the model runs, 20 degradation functions compute how well the specific configuration withstands the external stressors. The system subsequently calculates the seven key resilience metrics – reliability, availability, fault tolerance, integrity, recovery time, performance stability, and user safety – and aggregates them into a final fitness score. This score is utilised by the GA to iteratively refine the population, effectively breeding better architectural configurations over successive generations.
Two distinct architectural states were evaluated to provide a clear comparison of the method's effectiveness. First, a baseline architecture was established with minimal active mitigations to represent a standard, static AR/VR system design. This was compared against the synthesised, optimal architecture output by the Genetic Algorithm after convergence, representing a system "hardened" by the automated synthesis process. This comparative setup allowed for the quantification of the exact value added by the synthesis method, enabling us to observe how the algorithm traded off different design variables to survive the severe operational scenario.
The Genetic Algorithm successfully navigated the 20-dimensional design space to identify a configuration that significantly outperformed the baseline. The convergence of the algorithm is visualized in the penalty plot (Fig. 2), showing a steady improvement in the fitness function over 30 generations. The most significant finding of this experiment is the magnitude of improvement in time-critical metrics. Table 1 details the comparative scores.
Availability (+82.2%): This massive increase indicates that the synthesized architecture prioritized mechanisms like checkpoint/restart and redundant hardware paths. In a severe scenario where failures are frequent, the ability to keep the system "up" is the primary differentiator. Recovery Time (+39.9%): The optimization heavily favored strategies that reduce Mean Time To Repair (MTTR). By tuning software checkpoints and automated reset triggers, the system returns to a functional state much faster after a crash. Reliability (+31.1%): The increase in reliability suggests the effective use of thermal management and load balancing to prevent the occurrence of faults, rather than just managing them after they happen. The real-time behaviour of the optimized system (Fig. 3) showed that the analytical score successfully tracked changing environmental parameters, providing a responsive measure of resilience.
The synthesis framework presented in this study offers a significant methodological shift for systems engineers, moving the design process from manual, heuristic-based configuration to a calculated, automated optimization workflow. By translating qualitative mitigation strategies –such as dynamic voltage scaling, checkpoint-restart frequencies, or forward-error correction levels – into a quantifiable 20-dimensional mitigation vector, the method allows designers to mathematically evaluate the impact of architectural choices before physical prototyping. This is particularly valuable in the context of modern AR/VR systems, where the sheer complexity of the interaction between hardware, software, and user behaviour makes it nearly impossible for human designers to intuitively predict how a specific change (e.g., throttling GPU performance) will cascade through the system to affect end-user metrics like motion-to-photon latency or cybersickness.
Furthermore, the integration of this analytical block into a real-time Simulink environment demonstrates its utility as a dynamic runtime monitor, rather than just a static design tool. The experiment confirmed that the instantaneous analytical score successfully tracked changing environmental parameters in real-time. This capability suggests that the proposed model can be deployed as a "digital twin" or a runtime supervisor within the final product. In such a deployment, the system could dynamically adjust its own mitigation vector in response to detected environmental stress—automatically shifting from high-fidelity rendering to high-reliability modes when network jitter or thermal throttling is detected—thereby maintaining the aggregated resilience score above the minimal acceptance gates defined during the verification phase.
This paper has presented and experimentally validated a formally grounded method for the automated synthesis of resilient AR/VR architectures. By establishing a rigorous mathematical link between low-level architectural design choices and high-level emergent resilience properties, we transformed the complex, qualitative challenge of system design into a solvable multi-objective optimization problem. The experimental results provided a clear validation of this approach; the genetic algorithm successfully navigated a high-dimensional design space to identify a configuration that achieved a 19.4% improvement in the overall resilience score compared to a standard baseline. Most notably, the synthesis process prioritized critical operational metrics, yielding massive relative gains in Availability (+82.2%) and Recovery Time (+39.9%). These results confirm that automated synthesis can identify non-intuitive combinations of mitigation strategies that significantly enhance robustness without requiring manual trial-and-error.
Future research will focus on expanding the scope of this framework to address the increasingly hostile threat landscape facing interconnected immersive systems. While the current model effectively manages environmental and operational faults, the next iteration will incorporate defences against advanced security threats. Specifically, we aim to adapt the resilience model to account for malicious actors, including the impact of botnets and polymorphic malware on system integrity. This expanded focus will also address the unique privacy and security challenges inherent to immersive environments, such as protecting user authentication data and preventing non-immersive attacks that exploit the tight coupling between the user and the virtual environment. Additionally, we plan to refine the sensitivity models with empirical data gathered from physical testbeds and explore the application of this synthesis method to other distinct classes of cyber-physical systems.
Declaration on Generative AI The authors have not employed any Generative AI tools. [18] L. Di Geronimo, L. Braz, E. Fregnan, F. Palomba, and A. Bacchelli, “UI Dark Patterns and Where to Find Them: A Study on Mobile Applications and User Perception,” in Proc. CHI, 2020, pp. 1–14. doi: 10.1145/3313831.3376600. [19] A. Alshehri, K. Alrehili, F. Alhumaid, and A. Alessa, “Exploring the Privacy Concerns of Bystanders in Smart Homes,” Proc. Priv. Enhancing Technol. (PoPETs), vol. 2022, no. 4, pp. 253– 270, 2022. doi: 10.56553/popets-2022-0064. (Open PDF available.) [20] H. Lin and N. W. Bergmann, “IoT Privacy and Security Challenges for Smart Home
Environments,” Information, vol. 7, no. 3, art. 44, 2016. doi: 10.3390/info7030044. [21] Q. Wang, Z. Cai, R. Li, and X. Fang, “A Comprehensive Survey on Local Differential Privacy toward Data Statistics and Analysis,” Sensors, vol. 20, no. 24, art. 7030, 2020. doi: 10.3390/s20247030. [22] Y. Yao, T. Zimmermann, A. Chin, and F. Schaub, “Privacy Perceptions and Designs of Bystanders in Smart Homes,” Proc. ACM Hum.–Comput. Interact. (CSCW), vol. 3, no. CSCW, pp. 1–24, 2019. doi: 10.1145/3359161. [23] E. Saqib, S. He, J. Choy, R. Abu-Salma, J. Such, J. Bernd, and M. Javed, “Bystander Privacy in Smart Homes: A Systematic Review of Concerns and Solutions,” ACM Trans. Comput.–Hum.
Interact. (TOCHI), pp. 1–?, 2025. doi: 10.1145/3731755. [24] K. Marky, S. Prange, F. Krell, M. Mühlhäuser, and F. Alt, “‘You Just Can’t Know About Everything’: Privacy Perceptions of Smart Home Visitors,” in Proc. MUM, 2020, pp. 1–13. doi: 10.1145/3428361.3428464. (Open PDF available.) [25] A. P. Felt, E. Ha, S. Egelman, A. Haney, E. Chin, and D. Wagner, “Android Permissions: User Attention, Comprehension, and Behavior,” in Proc. SOUPS@CHI, 2012, pp. 3–14. doi: 10.1145/2335356.2335360. [26] V. Zimmermann, P. Gerber, K. Marky, L. Böck, and F. Kirchbuchner, “Assessing Users’ Privacy and Security Concerns of Smart Home Technologies,” i-com, vol. 18, no. 3, pp. 197–216, 2019. doi: 10.1515/icom-2019-0015. [27] O. Alshamsi, K. Shaalan, and U. Butt, “Towards Securing Smart Homes: A Systematic Literature Review of Malware Detection Techniques and Recommended Prevention Approach,” Information, vol. 15, no. 10, art. 631, 2024. doi: 10.3390/info15100631. [28] M. M. Ogonji, G. Okeyo, and J. M. Wafula, “A Survey on Privacy and Security of Internet of Things,” Computer Science Review, vol. 38, art. 100312, 2020. doi: 10.1016/j.cosrev.2020.100312.