Automating Concrete Simulation Scenario Generation for Autonomous Driving with Large Language Models

Engineering notes

The generation time is 3~4 seconds, which is significantly better than the 600~1500 seconds of the traditional method and the 10~15 seconds of the generalized large language model.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The need for high-quality simulation scenarios to verify the safety of autonomous driving systems is growing, but there are still obstacles to overcome, like the high cost and low efficiency of creating scenario files that satisfy simulation platform standards. To address the issues, this study suggests an automated approach for creating concrete autonomous driving simulation scenarios using a large language model. This approach enables the automated conversion of natural language input into standard scenario file output. The functional scenario generation stage uses the fine-tuned large language model for structured expression and improves the lightweight model deployment efficiency through knowledge distillation; the logical scenario generation stage involves mapping the standard parameter space and introducing constraint rules to ensure rationality; and the concrete scenario generation stage involves generating high-risk key parameters through data mining and generative adversarial networks to improve scenario realism and challenge. Validation on the CARLA platform demonstrates that the scenario files produced by this paper's method perform well in terms of structural integrity and semantic consistency. The generation time is 3~4 seconds, which is significantly better than the 600~1500 seconds of the traditional method and the 10~15 seconds of the generalized large language model. This demonstrates the efficiency advantage and engineering applicability of the method.

Links and sources

• Official / arXiv page

Comments