Efficient Motion Planning in Narrow Passages via Long-Short Edge Coupling and Stitching Sampling

Sampling-based motion planners often suffer from slow finite-time convergence in narrow passages and cluttered high-dimensional environments, where useful valid samples near critical bottlenecks are sparse and bidirectional search can become locally stagnant. Existing informed planners improve efficiency by restricting sampling to cost-relevant regions or by refining global graph approximations, but local collision-failure information and the structure of the current path are often underused when search progress stalls. This paper proposes Stitching Informed Trees (SIT*), an almost-surely asymptotically optimal informed bidirectional planner that augments the underlying informed bidirectional search framework with two local stitching mechanisms. Before an initial feasible solution is found, SIT* detects exhaustion of the current bidirectional search over the graph approximation, identifies a promising forward–reverse frontier pair, and constructs a local bridge-repair ellipsoid around the current search interface. After an initial feasible solution has been found, SIT* applies long-short edge coupling to partition the current path and construct chained local ellipsoids for path-aware refinement. These local regions concentrate sampling on crowded or detour-prone path segments where cost improvement is more likely, while the global informed bidirectional search process is preserved. Benchmark results on Narrow Passage and Random Rectangles problems from R² R 2 to R⁸ R 8 show that SIT* improves finite-time success rate, initial-solution convergence, and final solution cost compared with competitive informed-planning baselines.