SPACE: A Generalist Neural Solver for Symmetric and Asymmetric Routing Problems
Researchers propose SPACE (Spatial Pivot-Aligned Coordinate-free Embedding), a framework unifying symmetric and asymmetric vehicle routing problems (VRPs) within a single neural solver. Existing generalist solvers struggle with asymmetric settings due to input inconsistencies and structural differences. SPACE defines node positions via relative distances to pivots, using a bidirectional Fréchet representation with furthest pivot sampling for invariant embeddings. The method enables coordinate-free, pivot-aligned representations that work across both symmetric and asymmetric VRPs. The paper is published on arXiv (2605.24484v1).
Key facts
- SPACE unifies symmetric and asymmetric VRPs in a generalist neural solver.
- Existing solvers degrade in asymmetric settings due to input inconsistencies.
- SPACE uses spatial pivot-aligned coordinate-free embeddings.
- A bidirectional Fréchet representation with furthest pivot sampling is employed.
- The approach enables invariant node representations across problem settings.
- The paper is available on arXiv with ID 2605.24484v1.
- The method addresses real-world scenarios encompassing both symmetric and asymmetric problems.
- The framework is designed for generalist neural routing solvers.
Entities
Institutions
- arXiv