A meta-reasoning and dual-level cognitive decision-making approach for blind-spot safety in autonomous driving.

Complex traffic scenes with dynamic occlusions create hidden blind-spot hazards that challenge the reliability of autonomous driving decision control. This study aims to improve safety under such uncertainty by introducing a meta-reasoning-driven dual-level decision framework. The framework integrates object-level and meta-level processing to assess situational uncertainty and adapt decision responses accordingly, and incorporates a hippocampus-amygdala-inspired importance-sampling mechanism to enhance learning from error-prone cases. Systematic experiments show that the proposed method reduces collision risk and yields more reliable and efficient decision behavior compared with conventional approaches. Across multiple experimental settings, the proposed framework improves the success rate by approximately 2-18% and increases the driving index by about 4-23% relative to baseline methods. These results demonstrate that meta-reasoning can effectively strengthen safety-critical decision control in uncertain environments and provide a feasible direction for incorporating biologically inspired mechanisms into autonomous driving systems.