Realistic LiDAR Data Simulation for Autonomous Systems using Physics-Informed Learning

Accurate and realistic simulation of LiDAR data is critical for the development and validation of autonomous driving systems. However, existing simulation approaches often suffer from a significant sim-to-real gap due to oversimplified modelling of physical interactions and environmental factors. In this work, we present a physics-informed deep learning framework that bridges this gap by enhancing the realism of simulated LiDAR data using generative adversarial networks guided by domain-specific physical constraints for LiDAR intensity. Our method incorporates key physical factors such as range, surface material properties, angle of incidence, and environmental conditions along with their underlying physical relationships as constraints into the Cycle-Consistent GAN architecture, enabling it to learn realistic transformations from synthetic to real-world LiDAR intensity data without requiring paired samples. We demonstrate the effectiveness of our approach across multiple datasets, showing consistent improvements in statistical similarity metrics and downstream perception tasks such as semantic segmentation. The proposed algorithm has been integrated into the SimDaaS simulation engine, providing a robust tool for the research and industrial community to generate high-fidelity LiDAR data for training and evaluation of autonomous systems.