Reproduce “Hydra-Nav: Object Navigation via Adaptive Dual-Process Reasoning” 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
Zixuan Wang, Huang Fang, Shaoan Wang, Yuanfei Luo, Heng Dong, Wei Li, Yiming Gan
An object navigation agent from ByteDance Seed, Hydra-Nav, integrates an adaptive dual-process reasoning system within a single vision-language model. It achieves state-of-the-art success rates and improves search efficiency by learning to adaptively switch between deliberative reasoning and reactive execution, also demonstrating robust zero-shot transfer to a physical robot.
Key insights
- Unifying deliberative "slow" reasoning with reactive "fast" execution within a single VLM, guided by an adaptive switching mechanism, can substantially reduce computational overhead in embodied tasks.
- Strong temporal-spatial reasoning, informed by structured long-term memory, is essential for robust and efficient object navigation, particularly for long-horizon tasks.
- Iterative Rejection Fine-tuning (IRFT) provides a method for an agent to learn adaptive resource allocation, enabling it to trigger deep reasoning only at critical "stagnation points" detected during navigation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.