A Safe Self-evolution Algorithm for Autonomous Driving Based on Data-Driven Risk Quantification Model

TL;DR

This work tackles safe exploration in self-evolving autonomous driving by introducing a transformer-based data-driven risk quantification model (DD-RQ) that mimics human risk perception and yields a risk value $RQ$ and an importance ranking $IR$. The DD-RQ is tightly integrated with a Safe Self-Evolution controller whose adjustable safety limits modulate planning horizon $T_c$ and safe acceleration $a_{safe}$ to balance safety and learning progress. Key contributions include (1) a data-driven, risk-aware surrogate for safety evaluation, (2) an adjustable-safety framework that mitigates overly conservative behavior while preserving evolutionary potential, and (3) a virtual-real interaction platform validated in high-fidelity simulation and real-vehicle tests showing safe, efficient driving policies in dense and mixed traffic. The results demonstrate safe, reasonable actions across complex scenarios, reduced reliance on safety guards, and improved average speed without compromising collision-free operation, highlighting practical impact for safer autonomous learning systems.

Abstract

Autonomous driving systems with self-evolution capabilities have the potential to independently evolve in complex and open environments, allowing to handle more unknown scenarios. However, as a result of the safety-performance trade-off mechanism of evolutionary algorithms, it is difficult to ensure safe exploration without sacrificing the improvement ability. This problem is especially prominent in dynamic traffic scenarios. Therefore, this paper proposes a safe self-evolution algorithm for autonomous driving based on data-driven risk quantification model. Specifically, a risk quantification model based on the attention mechanism is proposed by modeling the way humans perceive risks during driving, with the idea of achieving safety situation estimation of the surrounding environment through a data-driven approach. To prevent the impact of over-conservative safety guarding policies on the self-evolution capability of the algorithm, a safety-evolutionary decision-control integration algorithm with adjustable safety limits is proposed, and the proposed risk quantization model is integrated into it. Simulation and real-vehicle experiments results illustrate the effectiveness of the proposed method. The results show that the proposed algorithm can generate safe and reasonable actions in a variety of complex scenarios and guarantee safety without losing the evolutionary potential of learning-based autonomous driving systems.

Figures (10)

• Figure 1: Overall architecture of safe self-evolution algorithm for autonomous driving based on data-driven risk quantification model.
• Figure 2: The mechanisms of human risk perception.
• Figure 3: Trajectory planning results for different planning times.
• Figure 4: Overall architecture of data-driven risk quantification model.
• Figure 5: Safe distance adjustment. When $\Delta S > 0$, the ego vehicle should slow down to adjust the safe distance, and when $\Delta S < 0$, the ego vehicle should accelerate to adjust the safe distance