ScholarWorks@숙명여자대학교

초록

End-to-end autonomous driving requires robust trajectory planning to ensure both safety and adaptability under diverse conditions. Traditional reinforcement learning (RL) and imitation learning (IL) approaches encounter challenges such as unstable training, poor generalization, and high sample complexity. For these reasons, flow matching (FM) in generative modeling is proposed as a promising alternative for action planning. Therefore, this paper proposes the feasibility of FM-based planning using a simplified environment as a controlled abstraction of road networks. The evaluation demonstrates that FM achieves rapid and stable convergence and generates near-optimal trajectories. Furthermore, the proposed algorithm produces smoother paths than the RL method. These results highlight the potential of FM to ensure both optimality and adaptability in data-driven generative models. Therefore, this paper establishes FM as a foundation for advancing end-to-end autonomous driving systems.

키워드