End-to-end models for autonomous driving aim to jointly optimize perception, motion prediction, and planning, enabling downstream components to exploit rich sensory information while reducing the impact of upstream perception errors. The primary motivation is to enable decoupled and scalable pretraining while preserving the adaptability and rich information flow characteristic of end-to-end models. In particular, scalability is limited by the combination of large perception architectures and the data-intensive nature of motion prediction and planning, which require substantial scenario diversity to accurately model complex agent interactions. However, how to effectively exploit the knowledge learned by these models in real-world, end-to-end settings remains underexplored problem. * Design and implement a discrete, structured interface that compresses and semantically organizes ...
mehr