Scaffolding Foundation Models into Physical-World Agents Pushes the Frontier of Long-Horizon Navigation
NavMCP links a VLM planner to a navigation model so long tasks can carry memory across repeated physical searches.
The framework has the VLM decide what evidence to seek, where to look, and when to stop, while the navigation foundation model executes each semantic sub-goal in closed loop. Its collaboration channels turn navigation rollouts into accumulated evidence, negative findings, and unresolved goals without retraining either model. The paper reports state-of-the-art results on HM-EQA, MT-HM3D, and EXPRESS-Bench, plus a 14.9-point gain over an episodic interface on HM-EQA. On a Unitree Go2, NavMCP reached 78.3% success, with a larger baseline margin as task horizon increased. HF Daily Papers' note
The framework has the VLM decide what evidence to seek, where to look, and when to stop, while the navigation foundation model executes each semantic sub-goal in closed loop. Its collaboration channels turn navigation rollouts into accumulated evidence, negative findings, and unresolved goals without retraining either model. The paper reports state-of-the-art results on HM-EQA, MT-HM3D, and EXPRESS-Bench, plus a 14.9-point gain over an episodic interface on HM-EQA. On a Unitree Go2, NavMCP reached 78.3% success, with a larger baseline margin as task horizon increased. HF Daily Papers' note
score 5