SwiftVLN for SatNav — Temporal-Biased History Sampling
Related Repositories
- SwiftVLN: training and evaluation code for these checkpoints.
- SatNav: satellite-image navigation environments, datasets, and evaluation tools.
This checkpoint is designed for SwiftVLN on SatNav, the continuous-state vision-and-language navigation benchmark over satellite imagery. It biases history-frame sampling toward recent observations while retaining coverage of the full trajectory history.
Model
- Starting checkpoint:
Qwen/Qwen2.5-VL-3B-Instruct - Training data: SatNav-v0.1 offline expert trajectories
- Training: 1 epoch, full-parameter fine-tuning, learning rate
2e-5 - Context: 32 RGB frames, up to 8 temporally biased history frames
- Prediction horizon: 4 actions
- Memory: per-frame average pooling with stride 2, sampling base 2.0
Repository name
swiftvln-satnav: SwiftVLN trained and evaluated on SatNav3b: Qwen2.5-VL 3B backbone1ep: trained for one epochf32s4: uses a 32-frame window and predicts four actionsoverlap0: uses non-overlapping training windowspf-h8-b2.0: samples up to eight history frames with temporal-bias base 2.0pool-s2: applies average pooling with stride 2noembed: uses no embedding enhancement
Reference Results
Results reported in the SatNav paper, memory-design ablation, row Temporal-biased sampling.
SR, SPL, and OS are percentages. Changes in SR and OS are percentage points relative to the SwiftVLN reference model.
| Split | SR (%) | SPL (%) | OS (%) | Delta SR (pp) | Delta OS (pp) |
|---|---|---|---|---|---|
Test Seen (val_seen) | 66.9 | 66.4 | 75.8 | +1.1 | +3.4 |
Test Unseen (val_unseen) | 55.8 | 55.3 | 66.3 | +2.1 | +2.2 |
Usage
Use this checkpoint with the SwiftVLN evaluation guide. Keep the full repository name unchanged because SwiftVLN derives the evaluation configuration from it.
This checkpoint is a history-sampling ablation in the SwiftVLN SatNav Ablation Model Zoo.