An Anytime Trajectory Optimizer for Accurately Parking an Autonomous Vehicle in Tiny Spaces

Abstract

Parking in tiny and cluttered slots is a critical challenge for automated parking systems, where the planning module serves as a core component responsible for generating collision-free and kinematically feasible trajectories. However, existing methods face significant limitations, such as geometric-based approaches generating paths rather than trajectories, mass point models failing in tiny berths, and full-dimensional models with large-scale constraints lead to time-intensive computations. To address these issues, this paper proposes an anytime trajectory optimizer for accurately parking an autonomous vehicle in tiny spaces cluttered with obstacles. The proposed method incorporates differentiable collision-avoidance constraints and a trust region strategy to improve computational efficiency. It can also iteratively refine the solution, enabling the optimizer to converge to a new minimum by breaking previous constraints and establishing new ones. Extensive simulations show the proposed method outperforms mainstream planners, effectively handling cluttered environments and tiny parking slots where others fail, or outperforming them in speed when they succeed. It reduces runtime by 90.19% compared to the baseline, while ensuring feasibility and continuous optimization.