Instance Image Navigation on InsImageNav
35.6Success Rate (SR)CompassNav (SFT+RFT)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CompassNav (SFT+RFT)Model Source=Our Models, Training Protocol=SFT+RFT, Backbone=Qwen2.5-VL-7B2025.10 | 35.6 | 14.8 | |
| GPT-o4-miniModel Source=Closed-Source Models, Model Scale=mini2025.10 | 33.4 | 13.2 | |
| GPT-4oModel Source=Closed-Source Models2025.10 | 29.8 | 13.2 | |
| Gemini-2.5-FlashModel Source=Closed-Source Models2025.10 | 24.1 | 9.7 | |
| CompassNav (SFT)Model Source=Our Models, Training Protocol=SFT, Backbone=Qwen2.5-VL-7B2025.10 | 23.9 | 7.9 | |
| Base ModelModel Source=Our Models, Backbone=Qwen2.5-VL-7B2025.10 | 19.8 | 5.2 | |
| Qwen2.5-VL-3BModel Source=Open-source Models, Model Scale=3B2025.10 | 8.7 | 3.2 | |
| Qwen2-VL-7BModel Source=Open-source Models, Model Scale=7B2025.10 | 5.3 | 3.6 | |
| LLama3.2-11BModel Source=Open-source Models, Model Scale=11B2025.10 | 3 | 2.5 |