模仿学习之外，端到端轨迹如何优化？轻舟一篇刷榜的工作......

Core Insights - The article discusses the development of CATG, a new trajectory generation framework based on flow matching, which addresses limitations in existing end-to-end autonomous driving systems [1][4][22] - CATG achieved a score of 51.31 in the NAVSIM V2 challenge, demonstrating its effectiveness in trajectory planning and robustness against out-of-distribution data [4][22] Background Review - End-to-end multimodal planning has become a key method in autonomous driving, significantly improving robustness and adaptability compared to single trajectory prediction methods [3] - Current multimodal methods often rely on imitation learning, leading to a lack of behavioral diversity due to insufficient strategy diversity in real trajectories [3][6] - Various alternative strategies have been proposed to capture a broader distribution of reasonable trajectories, but many still struggle with integrating safety constraints directly into the generation process [3][6] Proposed Framework - CATG completely abandons imitation learning and supports the flexible injection of explicit constraints during the generation process [4][22] - The framework integrates feasibility and safety constraints into the generation process through a progressive mechanism, utilizing prior perception anchor points [7][22] - CATG allows for controllable trade-offs between aggressive and conservative driving styles by using environmental reward signals as conditional inputs [7][13] Experimental Results - CATG was extensively evaluated in the NAVSIM V2 challenge, showcasing superior planning accuracy and robust generalization capabilities [4][14] - The model's training involved two phases: the first focused on training the flow matching process, and the second on fine-tuning the energy matching process [18][22] - The results indicated high compliance with various metrics, including 100% drivable area compliance and 98.21% no-at-fault collisions in stage one [19] Limitations - The computational cost of generating trajectories through 100-step sampling remains high, and accelerating the sampling process may compromise trajectory quality [21] Conclusion - The article concludes that CATG represents a significant advancement in end-to-end planning for autonomous driving, effectively incorporating flexible conditional signals and explicit constraints during trajectory generation [22]