A Semantic Observer Layer for Autonomous Vehicles: Pre-Deployment Feasibility Study of VLMs for Low-Latency Anomaly Detection

TL;DR

This paper introduces a semantic observer layer using a quantized vision-language model to detect anomalies in autonomous vehicles at low latency, ensuring safety during deployment.

Contribution

It demonstrates the feasibility of deploying a low-latency, quantized VLM-based semantic observer for anomaly detection in autonomous vehicles.

Findings

Achieved ~500 ms inference time with quantization and FlashAttention2.

Identified NF4 recall collapse as a deployment constraint.

Mapped performance metrics to safety goals through hazard analysis.

Abstract

Semantic anomalies-context-dependent hazards that pixel-level detectors cannot reason about-pose a critical safety risk in autonomous driving. We propose a \emph{semantic observer layer}: a quantized vision-language model (VLM) running at 1--2\,Hz alongside the primary AV control loop, monitoring for semantic edge cases, and triggering fail-safe handoffs when detected. Using Nvidia Cosmos-Reason1-7B with NVFP4 quantization and FlashAttention2, we achieve ~500 ms inference a ~50x speedup over the unoptimized FP16 baseline (no quantization, standard PyTorch attention) on the same hardware--satisfying the observer timing budget. We benchmark accuracy, latency, and quantization behavior in static and video conditions, identify NF4 recall collapse (10.6%) as a hard deployment constraint, and a hazard analysis mapping performance metrics to safety goals. The results establish a pre-deployment feasibility case for the semantic observer architecture on embodied-AI AV platforms.