This thesis explores how modern deep learning methods and neural scene representations can enable scalable, data-driven sensor simulation for autonomous driving. Instead of relying purely on manually engineered sensor models, the project investigates how reconstructed 3D/4D scene representations can serve as compact, editable scene priors from which realistic sensor observations can be generated.
At the core of the thesis is the use of Gaussian-based scene representations as a bridge between real-world recordings and generative sensor imitation. These representations capture the structure, appearance, and dynamics of driving environments and provide a flexible foundation for learning sensor-specific generation modules. By training on paired recordings from cameras and target sensors, the system should learn to imitate the observation behavior of different sensing modalities and apply these learned models to new reconstructed scenes
The work is positioned at the intersection of deep learning, neural rendering, generative AI, autonomous-driving simulation, and multimodal sensor modeling. Possible research directions include designing sensor-aware scene queries, building generative adapters for different sensor modalities, evaluating realism and temporal consistency, and studying how learned sensor imitation can support synthetic data generation for perception training and validation.
