transfuser

Why we included this project

TransFuser is a well-organized reference implementation for end-to-end driving models, bundling the full pipeline from data generation to evaluation. The architecture fuses camera images, LiDAR point clouds, and depth or segmentation inputs through a transformer to output driving actions, trained with imitation learning against a privileged autopilot in CARLA. The repo includes a dataset downloader, scripts for generating training data, multi-GPU training code, and an evaluation agent for the Longest6 benchmark, so it works as a complete starting point rather than just a model snippet. It is research code tied to CVPR 2021 and PAMI 2023 papers, so expect to handle CARLA setup and dependency pinning yourself. For teams exploring sensor fusion or imitation-based autonomy, comparing it with the group's PlanT and KING repos shows how these driving agents evolve.