GG-Nav

This repository contains the model weights for GG-Nav: Goal-oriented Grounding Chain-of-Thought Elicits Reasoning in Navigation Foundation Models.

Model Description

GG-Nav is a goal-oriented navigation model built upon Qwen3-VL-4B-Instruct. It learns a structured Goal-oriented Grounding Chain-of-Thought (CoT) through supervised fine-tuning to better exploit streaming visual observations, explicitly localize target instances, and support reliable navigation decisions.

GG-Nav supports two goal-oriented navigation tasks:

• Object-Goal Navigation (ObjectNav): Find and navigate to any instance of a specified object category.
• Interactive Instance-Goal Navigation (IIGN): Find a particular object instance among potentially ambiguous candidates, asking clarification questions when necessary.

The model employs a two-stage reasoning mechanism:

1. Observation-to-Evidence: Extracts goal-relevant information from intermediate frames between adjacent decision steps, producing bounding boxes, temporally consistent instance IDs, and summaries of accumulated evidence.
2. Evidence-to-Decision: Combines the current observation with accumulated evidence to ground target instances and generate navigation actions (view adjustments, pixel goals, or STOP) or clarification questions.

GG-Nav is fine-tuned using approximately 30K grounding-annotated videos and 60K navigation trajectories. It achieves 72.3% success rate on HM3D ObjectNav and 42.0% on VL-LN, improving over the strongest compared baselines by 3.1 and 16.6 percentage points, respectively. On CoIN (Seen Synonyms), it reaches 31.86% success rate, an improvement of 5.3 percentage points. We further demonstrate zero-shot transfer to unseen real-world environments using a Unitree Go2 robot.

Resources

• Paper: arXiv:2610.14541
• Project Page: GG-Nav Project Page
• Code: GitHub - DuangZhu/GG_Nav

Usage

For model setup, inference, and evaluation on HM3D ObjectNav, VL-LN, and CoIN, please refer to the GG-Nav GitHub repository. The repository provides environment setup instructions, data preparation details, and evaluation scripts.

Citation

If you find GG-Nav helpful in your research, please cite:

@misc{zhu2026ggnav,
  title={GG-Nav: Goal-oriented Grounding Chain-of-Thought Elicits Reasoning in Navigation Foundation Models},
  author={Shaohao Zhu and Wensi Huang and Siqi Zhang and Yuqiang Yang and Meng Wei and Chenyang Wan and Tianle Liu and Tai Wang and Jinming Xu},
  year={2026},
  eprint={2610.14541},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.14541}
}