Real-Time Visual Anomaly Detection in High-Speed Motorsport: An Entropy-Driven Hybrid Retrieval- and Cache-Augmented Architecture

Motorsport’s upcoming 2027 technical constraints reduce the role of active mechanical stabilizers and shift a larger share of vehicle-dynamics understanding to real-time perception and software. This paper introduces Agentic Visual Telemetry, a hybrid Retrieval-Augmented Generation (RAG) and Cache-Augmented Generation (CAG) framework designed to diagnose high-frequency dynamic regimes from onboard video under millisecond-level latency and edge-hardware limits. The approach combines (i) spatiotemporal gating to detect novelty and uncertainty, (ii) cache-first inference to reuse stable visual priors at O(1) cost, and (iii) safety-aware supervision with fail-silent operation and a safe-mode degradation strategy when thermal or compute margins shrink. We validate the framework on the Aspar-Synth-10K dataset, focusing on safety-critical phenomena such as suspension chatter. Retrieval grounding yields large gains over a memoryless baseline, improving Macro-F1 from 0.62 (B0) to 0.88 (B5), while maintaining real-time feasibility; a RAG-only oracle provides slightly higher PR-AUC but violates the latency envelope. Full precision–recall curves show that the proposed hybrid model preserves performance in the high-recall operating region for chatter detection, reducing false negatives consistent with the grounding hypothesis. Overall, the results demonstrate that high-fidelity video interpretation can be achieved within strict real-time constraints through cache-first, retrieval-grounded agentic perception, enabling robust visual telemetry for next-generation motorsport analytics.