SeeNav-Agent: Enhancing Vision-Language Navigation with Visual Prompt and Step-Level Policy Optimization

We propose SeeNav-Agent, a novel LVLM-based embodied navigation framework that includes a zero-shot dual-view visual prompt technique for the input side and an efficient RFT algorithm named SRGPO for post-training.

🔥 Updates

• 2025/12/3: Release our paper and model checkpoints.

🚀 Highlights

• 🚫 Zero-Shot Visual Prompt: No extra training for performance improvement with visual prompt.

• 🗲 Efficient Step-Level Advantage Calculation: Step-Level groups are randomly sampled from the entire batch.

• 📈 Significant Gains: +20.0pp (GPT4.1+VP) and +5.6pp (Qwen2.5-VL-3B+VP+SRGPO) improvements on EmbodiedBench-Navigation.

📖 Summary

• 🎨 Dual-View Visual Prompt: We apply visual prompt techniques directly on the input dual-view image to reduce the visual hallucination.

• 🔁 Step Reward Group Policy Optimization (SRGPO): By defining a state-independent verifiable process reward function, we achieve efficient step-level random grouping and advantage estimation.

📋 Results on EmbodiedBench-Navigation

📝 Main Results

🖌️ Training Curves for RFT

🖍️ Testing Curves for OOD-Scenes

📦 Checkpoint

🛠️ Usage

Setup

1. Setup a seperate environment for evaluation according to: EmbodiedBench-Nav and Qwen3-VL to support Qwen2.5-VL-3B-Instruct.

2. Clone the evaluation environment to a seperate training environment by:

conda create --name <your_env_for_train> --clone <your_env_for_eval> 
conda activate <your_env_for_train>

3. Further setup the seperate training environment according to: verl-agent and also Qwen3-VL to support Qwen2.5-VL-3B-Instruct.

Evaluation

Use the following command to evaluate the model on EmbodiedBench:

conda activate <your_env_for_eval>
cd SeeNav
python testEBNav.py

HINT: you need to first set your endpoint, API-key and api_version in SeeNav/planner/models/remote_model.py

Training

verl-agent/examples/srgpo_trainer contains example scripts for SRGPO-based training on EmbodiedBench-Navigation.

1. Modify run_ebnav.sh according to your setup.

2. Run the following command:

conda activate <your_env_for_train>
cd verl-agent
bash examples/srgpo_trainer/run_ebnav.sh

📚 Citation

If you find this work helpful in your research, please consider citing:

@article{wang2025seenav,
  title={SeeNav-Agent: Enhancing Vision-Language Navigation with Visual Prompt and Step-Level Policy Optimization},
  author={Zhengcheng Wang and Zichuan Lin and Yijun Yang and Haobo Fu and Deheng Ye},
  journal={arXiv preprint arXiv:2512.02631},
  year={2025}
}

Acknowledgements

Our code is built with reference to the code of the following projects: verl-agent and EmbodiedBench.