Building a fully trainable AD stack is a promising approach to efficiently build and scale an AD system. Transformer neural networks using tokenization of diverse information into a common representation and attention mechanisms to efficiently combine data in time and space foster understandable and flexible end-to-end architectures. An enabler to follow the fast-evolving research in end-to-end is a toolchain allowing quick application of such architectures from training over simulation into closed-loop real-world vehicles. The presentation gives an overview of the toolchain and presents insights gained from tests in simulation and closed-loop applications in the real world.

• An overview of the end-to-end state of the art
• Why transformer neural networks have a huge impact on end-to-end
• An example of a toolchain fostering rapid end-to-end experiments
• Challenged and blocking points getting end-to-end architectures into a vehicle
• Insights into first closed-loop experiments with a closed-loop end-to-end AD stack