General-purpose service robots need navigation systems capable of handling a wide range of human requests in unfamiliar environments, combining task generality with scene generality. Existing approaches often fine-tune multimodal large language models (MLLMs) to predict navigation actions, which ties their behavior to the coverage of navigation training data and can limit generalization to new requests and environments.
To address this, researchers propose SuperNav, an agentic navigation system built on a key insight: let the MLLM focus on interpreting requests, understanding scenes, and making decisions while preserving its general-purpose capabilities, and delegate motion execution to navigation tools.
SuperNav equips a pretrained MLLM with a specialized agent harness without navigation-specific fine-tuning. The harness supports decision-making through Navigation Skills, agent-oriented Tools for physical interaction, and task-progress and context management. A unified visual-point interface connects decision-making to motion by allowing the model to specify destinations directly in images and revise its decisions based on execution feedback.
According to the authors, these components together support sustained navigation across different task requirements and environments. SuperNav outperforms four evaluated baselines on instance-level, multi-object, and demand-driven tasks. Category-level evaluation on HM3D and deployment on a real quadruped robot further demonstrate its applicability across environments.
The system targets diverse tasks and scenes, from finding objects and visiting multiple targets to fulfilling high-level requests. A project page is available at https://zju3dv.github.io/SuperNav/.