Enhanced SIRRT*: Structure-Aware RRT* With Hybrid Smoothing and Bidirectional Rewiring

초록

Sampling-based motion planners such as Rapidly-exploring Random Tree* (RRT*) are widely used for optimal motion planning but often suffer from slow convergence and high variance across runs due to random initialization. Skeletonization-Informed RRT* (SIRRT*), which combines deterministic environment structure with sampling-based optimization, improves reliability by providing an initial solution with low variance across runs. However, the resulting path is often geometrically irregular, which limits the efficiency of subsequent optimization and the quality of the final path. This paper presents Enhanced SIRRT* (E-SIRRT*), a framework that addresses this limitation by introducing two refinements applied before the informed optimization begins: hybrid path smoothing and bidirectional rewiring. The smoothing module uses spline fitting and collision-aware correction to generate a geometrically smoother, collision-free initial path on an occupancy grid map. Subsequently, bidirectional rewiring reshapes the tree and its cost-to-come values around this refined path to enable more effective cost propagation. These refinements build on established smoothing and rewiring principles; the contribution lies in their structure-aware integration with SIRRT*'s free-space-skeleton-based deterministic initialization to maintain consistency between the refined path and the underlying tree. Across three 2-D benchmark environments, E-SIRRT* reduces the deterministic initial path cost by 12.3-34.5% compared with SIRRT* while preserving zero-variance initialization. In the real-world LiDAR-map benchmark, E-SIRRT* achieves 1.36% and 1.68% lower final path cost than Batch Informed Trees (BIT*) and Advanced BIT* (ABIT*), respectively, under a fixed time limit. Real-robot experiments further validate the executability of the resulting paths.

키워드