DeepPathNet: A Clearance-Aware Graph Neural Planner for Collision-Free Robotic Arm Motion in Narrow Passages

• Clearance-aware GNN encodes link-obstacle distances as edge features • Transformer decoder generates variable-length collision-free waypoints • Three-tier evaluation decomposes neural, refinement, and fallback roles • Largest gains (+7.4pp) on tightest passages with clearance below • 04m • Real-robot validation: 81.7 Planning collision-free motions for articulated manipulators in narrow passages remains challenging: classical samplers require many iterations to resolve tight geometric constraints, while learning-based planners operating on raw point clouds lack explicit reasoning about link–obstacle clearances. We present DeepPathNet, a planning framework that constructs a configuration-dependent heterogeneous scene graph encoding robot link keypoints and obstacle primitives as nodes, with edge-level clearance features capturing signed distances and surface normals between each link–obstacle pair. A clearance-aware graph neural network propagates these geometric signals through message passing, and a Transformer decoder autoregressively generates variable-length waypoint sequences. A collision-aware refinement module ensures safety via L-BFGS smoothing, exact collision checking, local repair, and classical fallback. We propose a three-tier evaluation protocol (Raw / Refined / Full) that transparently decomposes system contributions. Experiments on MotionBenchMaker with Franka Panda (7-DoF) and UR5 (6-DoF) show that DeepPathNet achieves 91.2% success on in-distribution scenes and 78.5% on narrow passages, with statistically significant improvements over MPiNets ( p < 0.05) and the largest gains on the tightest gaps ( + 7.4 percentage points for clearance < 0.04 m). Ablation studies, cross-robot transfer, noise robustness, and simulation-based scenario experiments (5 scenario types, 120 trials) validate the approach.