Overview of Hydra-Nav
Key Features of Hydra-Nav:
- 🌟 Adaptive Dual-Process Reasoning. Hydra-Nav is a unified VLM architecture that mimics human cognition by adaptively switching between a deliberative "slow system" (for analyzing exploration history and high-level planning) and a reactive "fast system" (for efficient low-level execution), balancing high success rates with computational efficiency.
- 🌟 Three-Stage Curriculum Learning. We propose a progressive training pipeline: (i) spatial-action alignment to strengthen trajectory planning; (ii) memory-reasoning integration to enhance temporal-spatial reasoning over long horizons; and (iii) iterative rejection fine-tuning (IRFT) to enable the agent to autonomously learn when to trigger for reasoning.
- 🌟 SOTA Performance & Efficiency. Hydra-Nav achieves state-of-the-art results on HM3D, MP3D, and OVON benchmarks, outperforming second-best methods by 11.1%, 17.4%, and 21.2% respectively. It also demonstrates significantly improved search efficiency under the new SOT (Success weighted by Operation Time) metric.
Architecture of Hydra-Nav
As illustrated in Figure 1, Hydra-Nav unifies high-level planning and low-level control within a single VLM architecture.
- 🐢 Slow System (Deliberative Reasoning): The cycle initiates here. The model analyzes the global history and constructs a structured long-term memory. It generates a detailed Chain-of-Thought (CoT) reasoning to formulate a high-level navigation plan.
- 🐇 Fast System (Reactive Execution): Responsible for efficient execution. Crucially, it is guided by the reasoning generated by the slow system (injected as the system prompt). It operates as a multi-turn dialogue to autoregressively decode low level meta actions that strictly follow the high-level plan.
-
🔄 Adaptive Switching:
The transition is self-triggered. When the agent completes a sub-goal or the current observation invalidates the existing plan (e.g., getting stuck), it outputs a special
obstoken to trigger the slow system for reasoning.
Evaluation Results
We evaluate Hydra-Nav on three standard benchmarks: HM3D and MP3D for object goal navigation, and OVON for open-vocabulary object navigatiosks. Following standard protocols, we report performance using success Rate (SR) and Success weighted by Path Length (SPL).
Table 1. Comparison with state-of-the-art methods on HM3D, MP3D, and OVON benchmarks. The attributes denote: Low: low-level action output, High: high-level planning/action output, Dual: dual-system architecture. RGB: uses RGB observations. Depth: uses Depth observations. The best and second best results are denoted by bold and underline.
Figure 2. Performance analysis of multi-turn IRFT across different benchmarks.
Figure 3. Reasoning Frequency analysis of multi-turn IRFT across different.
Results
- 🌟 Robust Zero-Shot Generalization. Hydra-Nav achieves a +21.1% gain on the challenging OVON Val-Unseen benchmark, effectively leveraging VLM knowledge and reasoning capability to handle unseen objects without prior training.
- 🌟 Effective Curriculum Learning. Hydra-Nav validates a three-stage pipeline (Base → SFT → IRFT) that integrates temporal-spatial reasoning and memory, boosting performance by 33.1% on HM3D during the SFT stage.
- 🌟 Cost-Efficient Adaptive Reasoning. Unlike existing VLM agents that trigger reasoning at every step (dense reasoning), Hydra-Nav activates "slow thinking" only at critical stagnation points. This selective mechanism reduces reasoning frequency to just 3.0% of per-step baselines while doubling the SOT.
BibTeX
@article{wang2026hydranav,
title={Hydra-Nav: Object Navigation via Adaptive Dual-Process Reasoning},
author={Wang, Zixuan and Fang, Huang and Wang, Shaoan and Luo, Yuanfei and Dong, Heng and Li, Wei and Gan, Yiming},
year={2026},
journal={arXiv pre-print},
url={https://arxiv.org/abs/2602.09972}
}