GNN-DIP: Neural Corridor Selection for Decomposition-Based Motion Planning

A team of researchers including Peng Xie, Yanliang Huang, Wenyuan Wu, and Amr Alanwar has developed GNN-DIP, a breakthrough motion planning framework that dramatically improves robot navigation through complex, narrow environments. The system addresses a fundamental challenge in robotics: traditional sampling-based planners struggle in tight spaces because they rarely place samples in critical narrow passages, and even when they do, collision checking frequently fails. GNN-DIP solves this by integrating a Graph Neural Network (GNN) with a two-phase Decomposition-Informed Planner (DIP), using neural networks to predict optimal corridors through obstacle graphs while maintaining mathematical completeness guarantees.

The technical approach involves partitioning free space into convex cells, where every narrow passage is captured exactly as a cell boundary. The GNN component predicts portal scores on the cell adjacency graph to bias corridor search toward near-optimal regions, avoiding the combinatorial explosion of candidate corridors that typically bottlenecks decomposition-based planners. In 2D environments, the system uses Constrained Delaunay Triangulation with the Funnel algorithm for exact shortest paths, while 3D environments employ Slab convex decomposition with portal-face sampling for near-optimal evaluation.

Benchmark results demonstrate remarkable performance improvements across multiple challenging scenarios. In 2D narrow-passage tests, 3D bottleneck environments with up to 246 obstacles, and dynamic 2D settings, GNN-DIP consistently achieved 99-100% success rates while delivering speedups ranging from 2x to 280x compared to traditional sampling-based baselines like RRT* and PRM. This represents a significant advancement in making robots more reliable and efficient in real-world applications where navigation through cluttered spaces is essential.