This repository provides pretrained LiMo model weights from the paper:
Less is More 🍋: Scalable Visual Navigation from Limited Data
Project page: https://leggedrobotics.github.io/less-is-more/
Code repository: https://github.com/leggedrobotics/less-is-more
Overview
LiMo is a vision-based, goal-conditioned navigation model that predicts future trajectories from monocular RGB observations.
It uses a DINOv2 ViT backbone together with a transformer decoder to generate sequences of SE(2) poses.
The checkpoints in this repository correspond to the different training settings described in the paper.
Checkpoints
| File | Description |
|---|---|
limo_trained_on_D_aug.safetensors |
trained on dataset D_aug (D_tel + D_geo) |
limo_trained_on_D_geo.safetensors |
trained on dataset D_geo (planner labeled) |
limo_trained_on_D_tel.safetensors |
trained on dataset D_tel (teleop data) |
limo_with_side_cams.safetensors |
model using additional side camera inputs (trained on D_aug) |
Configuration files:
limo_config.json— standard front camera setuplimo_side_cams_config.json— multi-view (front + side cameras)