Reproduce “SuperNav: An Agentic Navigation System for Any Task in Any Scene” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinkai Zhang, Jingyi Xu, Yuanhong Yu, Jiarui Guo, Ruizhen Hu, Hujun Bao, Xiaowei Zhou, Sida Peng
SuperNav uses a pretrained multimodal large language model (MLLM) to navigate environments across single-object, ordered multi-object, and demand-driven tasks. A harness provides tools, reusable Skills, task state, and context management; the model selects destinations while an executor handles motion. On 150 Habitat-GS single-object tasks, its Geo-based Executor achieves 78.00% success versus UniNaVid's 34.00%, supporting the complete system only in tested settings.
Key insights
- Separating task interpretation from motion execution lets the system accommodate different requests and swap navigation backends through the same interface.
- Reusable guidance and multi-view observations materially affect performance: removing Skills or limiting the agent to a forward-facing view reduces success in the reported ablations.
- Success falls as ordered tasks grow longer, underscoring the difficulty of sustaining progress and completion judgments across many stages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.