Physics-Aware LiDAR Simulation

This project aims to:

1) Develop a physics-based simulation for the LiDAR sensor using deep learning on real-world data to decrease the discrepancy between the real LiDAR and currently common simulations.

2) Generate 3D assets automatically from the LiDAR point cloud instead of hiring a graphic designer to create the asset, so the generated assets are faster and cheaper to build. Also, these assets are used for a LiDAR simulation environment, so building them from LiDAR point clouds represents the real-world properties of LiDAR sensors much better. Note that these generated assets can be used for any LiDAR sensor, regardless of how they are created.

3) Generate critical safety scenarios using adversarial attacks on motion planning modules. These scenarios can help researchers examine the shortcomings of their motion planning models.

• Tech Stack: Python with PyTorch, Open3D, NumPy, Matplotlib, Pandas, scikit‑learn, Ubuntu, Git