Navigation on HM3D 2.5k episodes
33.4SCTDune
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DuneBackbone=Dune, Projection Head=Cross-attention pooling (CAP), Encoder=Frozen, Training Steps=500M2026.06 | 33.4 | 76.9 | 90.7 | |
| VC-1Backbone=VC-1, Projection Head=Cross-attention pooling (CAP), Encoder=Frozen, Training Steps=500M2026.06 | 28.9 | 72.2 | 86.3 | |
| AM-RadioBackbone=AM-Radio, Projection Head=Cross-attention pooling (CAP), Encoder=Frozen, Training Steps=500M2026.06 | 27.6 | 68.2 | 85.2 | |
| DinoV2Backbone=DinoV2, Projection Head=Cross-attention pooling (CAP), Encoder=Frozen, Training Steps=500M2026.06 | 16.1 | 49.8 | 63.4 | |
| *ResNetBackbone=*ResNet, Note=Half-width ResNet, Encoder=Trained from scratch, Training Steps=500M2026.06 | 15 | 41 | 53.5 | |
| DinoV3Backbone=DinoV3, Projection Head=Cross-attention pooling (CAP), Encoder=Frozen, Training Steps=500M2026.06 | 0 | — | — |