Embodied Navigation on GOAT-Bench
69.1Success Rate (SR)3D-Mem* (GPT-4o)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| 3D-Mem* (GPT-4o)Method Category=Closed-Source VLMs, Backbone=GPT-4o2026.05 | 69.1 | 48.9 | |
| SAGE (Qwen3-4B)Method Category=Open-Source VLMs, Backbone=Qwen3-4B2026.05 | 64.8 | 44.9 | |
| SAGE (Qwen3-2B) (Ours)Method Category=Open-Source VLMs, Backbone=Qwen3-2B2026.05 | 56.7 | 38.9 | |
| Explore-EQA*Method Category=Closed-Source VLMs2026.05 | 55 | 37.9 | |
| 3D-Mem* (LLaVA-7B)Method Category=Open-Source VLMs, Backbone=LLaVA-7B2026.05 | 49.6 | 29.4 | |
| 3D-Mem (Qwen3-2B)Method Category=Open-Source VLMs, Backbone=Qwen3-2B2026.05 | 46.4 | 20.3 | |
| SenseAct-NN Skill ChainMethod Category=RL Baselines2026.05 | 29.5 | 11.3 | |
| SenseAct-NN MonolithicMethod Category=RL Baselines2026.05 | 12.3 | 6.8 |