KAIST Visual Intelligence Lab

End-to-end autonomous driving learns a direct mapping from raw sensor inputs to driving actions using a single neural network or pipeline. By unifying perception, prediction, and planning into one model, it reduces system complexity and enables consistent, real-time decision-making. This approach forms a promising path toward scalable and robust driving intelligence. Vision foundation models are large-scale models trained on vast and diverse datasets, capable of adapting to many downstream visual tasks with minimal supervision. They support prompt-based interactions and generalize well across domains, reducing the need for task-specific training. These models provide a powerful backbone for modern vision systems across research and industry. Novel sensors provide unique forms of visual information beyond conventional RGB cameras, enabling perception in challenging conditions. Event cameras, 360° panoramic cameras, and other non-traditional sensors offer high temporal resolution, wide field of view, or high dynamic range. These sensors open new possibilities for robust perception in fast, bright, or complex environments. Event cameras asynchronously capture changes in brightness with microsecond latency, avoiding motion blur and handling extreme lighting. They are ideal for high-speed motion analysis, HDR imaging, and continuous-time perception. This sensor enables new algorithmic paradigms that differ fundamentally from frame-based vision. Multi-modal learning integrates complementary information from different sensor modalities to achieve more reliable and complete perception. By combining visual, geometric, semantic, and linguistic inputs, models can interpret complex real-world environments more effectively. This approach improves robustness under noise, occlusion, a