Simulating Realistic LiDAR Data Under Adverse Weather for Autonomous Vehicles: A Physics-Informed Learning Approach

Abstract

Accurate LiDAR simulation is crucial for autonomous driving, especially under adverse weather conditions. Existing methods struggle to capture the complex interactions between LiDAR signals and atmospheric phenomena, leading to unrealistic representations. This paper presents a physics-informed learning framework (PICWGAN) for generating realistic LiDAR data under adverse weather conditions. By integrating physicsdriven constraints for modeling signal attenuation and geometryconsistent degradations into a physics-informed learning pipeline, the proposed method reduces the sim-to-real gap. Evaluations on real-world datasets (CADC for snow, Boreas for rain) and the VoxelScape dataset show that our approach closely mimics realworld intensity patterns. Quantitative metrics, including MSE, SSIM, KL divergence, and Wasserstein distance, demonstrate statistically consistent intensity distributions. Additionally, models trained on data enhanced by our framework outperform baselines in downstream 3D object detection, achieving performance comparable to models trained on real-world data. These results highlight the effectiveness of the proposed approach in improving the realism of LiDAR data and enabling robust perception under adverse weather conditions.

Figures (8)

• Figure 1: Simulation-to-Real Gap: LiDAR point cloud with intensity information: a) Simulated Domain—VoxelScape data augmented with snow using a physics-based method lisa_kilic2021lidar; b) Real Domain—CADC data collected in snowy conditions. Objects marked with A (vegetation), B (metallic pole), C (car), and D (road surface). The significant difference in intensity distribution arises because the physics-based simulation fails to capture intensity attenuation due to adverse weather conditions like snow.
• Figure 2: Physical Modalities:(a) The range is calculated using the Euclidean distance between the LiDAR sensor and the point, while the incidence angle is determined using the dot product of the LiDAR ray direction vector and surface normal vectors. (b) Material reflectance is obtained by mapping semantic labels in the LiDAR data to spectral reflectance values from NASA's ECOSTRESS Library, based on identified materials.
• Figure 3: PICWGAN Training framework: LiDAR point clouds, projected onto a spherical surface, are input into our proposed PICWGAN architecture to generate LiDAR intensity under adverse weather conditions.
• Figure 4: Architecture: Generator and Discriminator architecture.
• Figure 5: Qualitative Analysis - 3D: Visual assessment of the back-projected 3D point cloud with intensity information by our PICWGAN architecture trained on the snow weather CADC dataset. (a) Physics-based simulated snow-weather LiDAR point cloud from the VoxelScape dataset; (b) Same point cloud (from a) after domain adaptation by our PICWGAN; (c) Ground truth LiDAR point cloud from the CADC dataset used for training the PICWGAN.