Local navigation target prediction on Landmark-RxR occluded-target
42.83Geodesic Accuracy (GeoAcc)BEACON
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| BEACONInput=RGB-D, Output=BEV heatmap2026.03 | 42.83 | 45.36 | 2.6 | 43.56 | |
| VLM + point headInput=RGB, Output=BEV point2026.03 | 32.15 | 39.17 | 20 | 36.99 | |
| RoboPoint-13B (best point)Input=RGB, Output=image point2026.03 | 29.14 | 37.72 | 26.96 | 36.42 | |
| RoboRefer-8B-SFTInput=RGB-D, Output=image point2026.03 | 20.09 | 25.45 | 21.49 | 23.65 | |
| RoboPoint-13BInput=RGB, Output=image point2026.03 | 15.43 | 23.46 | 35.18 | 21.88 | |
| ChatGPT-4o (oracle-view)Input=RGB, Output=image point2026.03 | 9.52 | 17.11 | 41.68 | 15.31 | |
| ChatGPT-4oInput=RGB, Output=image point2026.03 | 5.69 | 11.55 | 54.03 | 10.39 |