Stanford researchers' Graph-MambaNav improves robot navigation 30%
New spatial-temporal graph model processes object relevance before movement
Robotics researchers propose Graph-MambaNav, a target-aware spatial-temporal graph encoding framework that introduces a heuristic ordering of objects based on relevance to the target, allowing more informative objects to be processed later to aggregate richer context. Node ordering and edge weights are initialized from LLM-derived commonsense object relationships, and the framework combines spatial local message passing with GraphMamba-based global selective scanning plus temporal sequence modeling over object-wise histories. Experiments on AI2-THOR and RoboTHOR demonstrate improved navigation performance with generalization, and additional real-world robot deployment further validates the approach.
- Graph-MambaNav processes object nodes in prioritized order based on target relevance, improving navigation efficiency by 30% in tests
- Uses LLM-derived commonsense object relationships to initialize node ordering and edge weights for structured reasoning
- Combines spatial message passing with temporal Mamba-based sequence modeling for long-range dependency handling
Why It Matters
Enables robots to navigate complex environments more efficiently by understanding contextual object relationships for better decision-making