Real-time motion planning for autonomous parking on embedded advanced driver-assistance system (ADAS) platforms faces a fundamental computational bottleneck: transformer-based approaches (e.g., Motion Planning Trans-former, Diffusion-based planners) achieve strong performance but incur prohibitive computational costs unsuitable for resource-constrained automotive systems. This work proposes a lightweight alternative machine learning approach using Random Forest classifiers and regressors to predict parking trajectory regions and vehicle orientations, enabling accelerated Rapidly-exploring Random Trees (RRT) planning without sacrificing robustness. The approach is trained on a dataset of 10,725 synthetic per-pendicular backward parking scenarios generated via Rapidly-exploring Random Tree Star (RRT*) in the Reeds-Shepp con-figuration space. Using Random Forests with 20 trees and maximum depth 8, the method achieves 98.3–100% success rate in multi-direction-change scenarios with planning times of 0.15–0.25 seconds, compared to 2.81 seconds for unconstrained RRT. In scenarios with insufficient prediction guidance, the constrained planner can maintain a fallback mechanism that preserves RRT’s probabilistic completeness guarantees. This work demonstrates that simpler machine learning models can match transformer-based approaches while remaining practical for embedded deployment.
Kamel et al. (Thu,) studied this question.