SwiftVLN for SatNav — Relative Pose

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 augments visual tokens with relative pose through feature-wise linear modulation.

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 and up to 8 uniformly sampled history frames
• Prediction horizon: 4 actions
• Memory: per-frame average pooling with stride 2
• Pose fusion: FiLM over relative position and heading features

Repository name

• swiftvln-satnav: SwiftVLN trained and evaluated on SatNav
• 3b: Qwen2.5-VL 3B backbone
• 1ep: trained for one epoch
• f32s4: uses a 32-frame window and predicts four actions
• overlap0: uses non-overlapping training windows
• pf-h8-pool-s2: uses the reference per-frame memory configuration
• posefilm: injects relative pose into visual tokens with FiLM

Reference Results

Results reported in the SatNav paper, memory-design ablation, row Relative pose.

SR, SPL, and OS are percentages. Changes in SR and OS are percentage points relative to the SwiftVLN reference model.

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 an input-augmentation ablation in the SwiftVLN SatNav Ablation Model Zoo.