A neuromorphic safety monitor for verifiable runtime assurance in stochastic control loops
Year: 2026
Type: article
Abstract: The integration of high-performance but non-deterministic AI into safety-critical control loops remains a fundamental challenge for verifiable autonomy. We present a runtime assurance framework in the simplex family, in which a monitored primary controller is paired with a certified fallback safety controller. The monitor is a neuromorphic anomaly detector that we model as a stochastic sensor and characterize directly in this manuscript via end-to-end measurements of its detection latency distribution and its false negative and false positive rates. Using these measured quantities, we derive a probabilistic bound on closed-loop failure risk over a finite horizon that explicitly links monitor reliability and response time to safety. We validate the framework in a high-fidelity cislunar rover simulation under combined adversarial and environmental faults, where timely intervention yields an absolute mission-success gain of 84 percentage points. The result is an auditable methodology for deploying advanced AI in control systems where rapid and reliable fault detection is required.
Source: Neurocomputing
Author: Sylvester Kaczmarek
Institution: Imperial College London
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,670
OpenAlex ID: W7137124550