lauriasuo/EarthRovers_ViNT_nLoc32_nH32

PointGoal navigation policy trained with the FrodoBots8K dataset.

Presented in the paper Data scaling for navigation in unknown environments

Project page | Code

Details

Architecture:

ViNT-like, Theia encoder + transformer backbone + MLP head

Dataset:

1024H total, 32H from 32 distinct geographical locations

Usage

See the Github repo