NaVILA for SatNav — Navigation-pretrained initialization
This checkpoint is designed for use with SatNav, the continuous-state vision-and-language navigation platform for satellite maps. The SatNav NaVILA integration uses it to predict compact actions from a language instruction and eight RGB observations.
Model
- Starting checkpoint:
a8cheng/navila-llama3-8b-8f - Training data: SatNav-v0.1 offline trajectories
- Training: 1 epoch, full-parameter fine-tuning
- Action format:
stop,forward,left, andright
Repository name
navila-satnav: NaVILA adapted to SatNavcontinue: navigation-pretrained initialization from the NaVILA navigation SFT checkpoint1ep: trained for one epoch8f: uses eight visual frames per samplesample: the following fields describe trajectory-step samplinghk7: keeps the first seven action stepsfs7: keeps every seventh step in a consecutiveforwardrun after the head; turning steps are keptstopx4: repeats each terminalstopsample four times
Reference Results
Results reported in the SatNav paper, baseline comparison, row NaVILA (navigation-pretrained).
SR, SPL, and OS are percentages. NE is measured in meters. Steps is the average number of executed environment actions, including turns and stop.
| Split | SR (%) | SPL (%) | OS (%) | NE (m) | Steps |
|---|---|---|---|---|---|
Test Seen (val_seen) | 25.0 | 24.9 | 35.0 | 93.88 | 51.47 |
Test Unseen (val_unseen) | 18.6 | 18.4 | 31.6 | 123.49 | 59.57 |
Usage
Use this checkpoint with the SatNav NaVILA integration. It is also listed in the SatNav Baseline Model Zoo.
License
The checkpoint contains a Llama 3 backbone and is released under the applicable Llama 3 terms. NaVILA, SatNav, and other components retain their own licenses.