When Does Predicting Failure Help? A Boundary Analysis of Predictive versus Reactive Runtime Assurance for Multi-Robot Coordination
A prediction can tell a robot to act before a safety margin runs out. But afixed threshold can also make it act early. I wanted to understand whenprediction adds something beyond choosing a more conservative threshold. Istudy this question with a simplified timing model and two kinematic simulationsetups, giving both rules the same tuning budget and testing on separate noiseseeds. In the timing model, the predictive rule is a current-margin threshold whosebuffer scales with the estimated rate of decline. Extra lead changes therecovery outcome only within a band set by recovery duration and deadlineslack; old observations consume that lead one for one. The minimal predictivehorizon wastes less lead than the minimal safe fixed buffer when recovery needsa similar amount of time across cases but the corresponding margins differ,provided the rates can be estimated. The clearest gain appears in the three-robot shared-zone test. Both tunedrules succeed in all 1,620 conflict episodes, and prediction reduces meancompletion time by 2.78 s (paired 95% bootstrap interval [2.66,2.90] s). Onclear paths, it reduces unnecessary serializations from 1,080 to 21 out of1,620 episodes. The two-robot test exposes a different behavior: the tunedfixed rule records no collisions in 6,750 conflict episodes and predictionrecords two. Together, the tests show where adapting the margin helps in thesesimulations and where estimating its rate becomes the weak point.
Authors
- Mohamed Adil Farukh
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24242719
- Primary Topic
- Real-Time Systems Scheduling
- Type
- preprint