This research presents a machine learning pipeline for evaluating the realism of synthetic LiDAR point clouds generated by AWSIM, an open-source simulator developed by TIER IV for Autoware, a leading self-driving system. The LiDAR realism evaluation enables reliable simulation testing for Autoware's perception algorithms. Using geometric descriptors and XGBoost classification, we quantify how closely simulated LiDAR data resembles real-world sensor data. Our experiments validated that Distance Gaussian Noise significantly improves synthetic data realism, particularly on vertical surfaces. The pipeline achieves 91% accuracy and provides actionable feedback through SHAP-based feature importance analysis and realism score visualization.

• Practical Methodology : a pipeline approach to quantify LiDAR simulation realism using geometric descriptors and realism scoring metrics
• how real-world LiDAR data from TIER IV's actual autonomous driving operations enables reliable simulation realism evaluation
• how simulation validation supports Autoware systems that TIER IV has deployed in real-world services and experiments
• insights into AWSIM and Autoware simulation tools available for their own autonomous driving development projects
• key differences between synthetic and real LiDAR data and practical approaches to bridge this critical gap