Mixing real and synthetic driving data smarter, not just more

Training self-driving systems on real driving footage is expensive and limited; synthetic data is cheap but adding all of it hurts performance due to domain mismatch. AutoScale automatically decides which synthetic scenes to include during training by treating data selection as an optimization loop: it represents scenes using a graph-based autoencoder, estimates which clusters matter most, and retrieves high-value samples. On NavSim benchmarks, it achieves better results than naive real-synthetic mixing while using significantly fewer synthetic samples under fixed compute budgets.