SwiftVLN Documentation

SwiftVLN brings trajectory-based training, memory experiments, and online evaluation into one framework. It supports Qwen2.5-VL and Qwen3-VL for SatNav aerial navigation and Habitat indoor navigation.

Start here

Your goal	Guide
Understand the implementation	Dual memory and windows · History compression · Map memory · Input augmentation
Evaluate a released model	Installation · Checkpoints · Evaluation
Train a navigation policy	SatNav data · Habitat data · Training
Compare memory designs	Memory configuration
Adapt to UAV observations	SatDronePair · Stage-A training
Extend the framework	Architecture · Extension guide

Training and evaluation workflow

Prepare an environment, download a base model or navigation checkpoint, and configure local data paths. Training uses offline RGB observations and expert actions; evaluation runs the policy online on SatNav or Habitat episodes. Use the same model name to carry memory and window settings from training into evaluation.

The workflow guide connects the setup, data, training, evaluation, and development documents. The reference configuration uses 32-action windows, four actions per prediction, and eight uniformly sampled history frames.

Project resources

• Paper: SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery.

• SwiftVLN Model Zoo: reference and memory ablation checkpoints.

• SatNav Wiki: simulator, episodes, satellite scenes, and baseline documentation.

• Source code: models, scripts, configuration, and documentation sources.