EvoEye: Self-Evolving Runtime Monitoring for Autonomous Driving Systems

2026-07-04

AI Summary

Current runtime monitors for autonomous driving struggle with unforeseen risks and suffer from persistent prediction errors. This work proposes a self-evolving monitoring framework that integrates a collision prediction mechanism leveraging cross-module temporal interactions (FusionMonitor) with a density-aware mutation-based strategy for blind spot exploration (BlindSpotEvolver). By harnessing prediction errors to guide the generation of informative new scenarios, the framework iteratively updates its models to autonomously enhance monitoring capabilities. Experiments on the Apollo and CARLA platforms demonstrate that the approach improves frame-level recall by up to 37.8 percentage points, achieves median warning lead times of 2.8–4.2 seconds, and increases the F1 score for blind spot sampling by up to 13.2 points over baseline methods.

Abstract

Runtime monitoring is essential for detecting impending hazards in autonomous driving systems (ADSs). However, existing ADS runtime monitors have fixed detection capabilities: rule-based monitors cover only manually specified hazards, while learning-based monitors depend heavily on their initial training data and may retain substantial prediction errors. We therefore propose EvoEye, which identifies the current monitor's errors, generates informative executions accordingly, and updates the monitor through self-evolution. To enable effective self-evolution, EvoEye combines a capable runtime monitor with targeted scenario acquisition. FusionMonitor learns cross-module temporal interactions for collision prediction, while BlindSpotEvolver converts current prediction errors into search guidance and uses density-aware mutation to acquire informative executions for subsequent monitor updates. We evaluate EvoEye on Baidu Apollo with CARLA in representative highway and urban scenarios. FusionMonitor improves frame-level Recall by up to 37.8 percentage points at a false positive rate of 0.05, with 2.49 ms latency and 2.8-4.2 seconds of median warning time. Under the same budget, BlindSpotEvolver outperforms uniform and violation-oriented sampling by up to 13.2 F1 points on previously missed unsafe contexts.

Problem

• runtime monitoring
• autonomous driving systems
• detection capability
• prediction errors
• monitoring blind spots

Innovation

• self-evolving monitoring
• runtime monitoring
• autonomous driving systems
• density-aware mutation
• cross-module temporal interaction