Submitted:
15 November 2026
Posted:
18 November 2026
You are already at the latest version
Abstract
Reliability models can misrepresent redundant paths when communication links are treated as series elements. We tested this assumption through accelerated process-level fault injection in six running ROS 2 control architectures, comparing precomputed continuous-time Markov predictions based on graph logic (A) and a series-link rule (B). A pilot recorded 360 system failures; separate protocol-frozen studies recorded 960 signal-level failures, 400 trials with imposed common-cause failures and 540 trials with message loss, alongside 131 selected production windows from six simulated cells. No model parameter was fitted to observed failure times. In the two architectures that discriminate between A and B, the series-link model underestimated the observed mean time to signal loss by 44–52% and fell outside the prespecified confidence intervals. It also underestimated production in selected windows by 14–19%, although formal inference varied with the interval method and incident accounting. Model A met the prespecified agreement rule in five of six architectures at signal level; its production-rate errors were 1.7–3.8% under the prespecified summary, with sensitivity to window dependence. A statistically established advantage of parallel over single links was not obtained (difference 4.0 s, 95% CI −1.1 to 9.2). Under imposed common-cause failures, an independence model overestimated mean time to loss by up to 65%, while an injector-matched common-cause extension stayed within 9%. Agreement deteriorated under severe message loss. The findings support a falsification-oriented test of architecture-level reliability assumptions in running middleware, not claims about hardware failure rates or universal robotic reliability.
Keywords:
ROS 2
; fault injection
; reliability
; Markov chains
; redundancy
; industrial automation
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.