3D Trajectory Prediction on DroidSpatial-Bench
63.5Accuracy (Start, δ=5 cm)3D HAMSTER (Qwen3-VL-8B + Ldepth)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| 3D HAMSTER (Qwen3-VL-8B + Ldepth)Input=RGBD2026.06 | 63.5 | 66.2 | 41.9 | 80.4 | 82.4 | 65.5 | |
| 3D HAMSTER (Qwen3-VL-8B + Depth encoder)Input=RGBD2026.06 | 62.8 | 62.2 | 42.6 | 83.8 | 75 | 62.8 | |
| RoboBrain-2.5-8BInput=RGB2026.06 | 61.5 | 58.1 | 39.2 | 80.4 | 74.3 | 60.1 | |
| 3D HAMSTER (Qwen3-VL-8B + 3D Traj. Data)Input=RGB2026.06 | 50 | 50 | 27.7 | 71.6 | 72.3 | 50 | |
| Gemini-3.0-ProInput=RGB2026.06 | 29.1 | 44 | 16.2 | 43.2 | 56.1 | 29.7 | |
| GPT-5.2Input=RGB2026.06 | 6.8 | 29.7 | 2.7 | 29.7 | 45.3 | 16.2 | |
| Sonnet-4.6Input=RGB2026.06 | 1.4 | 8.8 | 0.7 | 6.1 | 16.2 | 2 | |
| Qwen3-VL-8BInput=RGB2026.06 | 0.7 | 9.5 | 0.7 | 0.7 | 14.9 | 0.7 |