Less is More 🍋: Scalable Visual Navigation
from Limited Data

📄 Paper | 🌐 Project Page | 🤗 Weights | 📊 Dataset

Less is More (LiMo) is a transformer-based visual navigation policy that predicts goal-conditioned SE(2) trajectories from a single RGB observation. We demonstrate that augmenting limited expert demonstrations with geometric planner-generated trajectories yields substantial performance improvements, achieving robust visual navigation through strategic data curation rather than simply collecting more data.

Release Status

• Inference code
• Checkpoints released (SafeTensors on HuggingFace)
• Dataset and training code
• Dataset builder with MPPI planner
• ROS integration

Installation

This project uses uv for dependency management. Tested on Ubuntu 22.04, Python 3.10, NVIDIA GPU with driver version ≥530 (CUDA 12.1).

git clone https://github.com/leggedrobotics/less-is-more && cd less-is-more && uv sync

To also use the dataset builder:

uv sync --extra dataset_builder

Note: The dataset_builder extra is only supported on Linux x86_64 with Python 3.10. It vendors a pre-built FastGeodis wheel because pip install FastGeodis requires a from-source CUDA build that is broken in the standard distribution.

The codebase uses Hydra for configuration management in the limo package.

Download Pretrained Weights

Download the pretrained LiMo checkpoint:

wget -O data/weights/limo_trained_on_D_aug.safetensors \
  https://huggingface.co/yv1es/less-is-more/resolve/main/limo_trained_on_D_aug.safetensors

Quick Start

Run inference on the provided examples:

uv run limo/src/inference.py

The default configuration in limo/configs/inference.yaml processes images from the example dataset in data/inference_example/.

Example: Inside ETH Zurich main building with multiple predictions

Example: Outdoor mountain trail with multiple predictions

Configuration

The inference pipeline is configured via limo/configs/inference.yaml:

• weights_path: Path to model weights (default: data/weights/limo_trained_on_D_aug.safetensors)
• input_path: Single image file or directory containing .jpg, .png, or .jpeg files
• goals_csv: CSV file defining navigation targets (format: x,y,yaw in meters/radians, robot frame)
• camera_info: Optional YAML file with camera intrinsics for path projection visualization
• output_dir: Directory for generated visualizations

See data/inference_example/ for a complete working example with sample images, goals, and camera calibration.

Training

The training pipeline uses PyTorch Lightning for model training and Hydra for configuration management.

Training Script

Train the model using: :

uv run limo/src/train.py experiment=train_limo_debug

(this debug experiment config will only start a quick debug run to test your system)

The main entry point is limo/src/train.py, which:

• Loads configuration from limo/configs/train.yaml and an experiment config
• Automatically pulls the dataset from HuggingFace
• Runs training with logging
• Saves checkpoints and weights

Configuration System

The training pipeline uses Hydra with the following structure:

Config directories (limo/configs/):

• data/: Dataset configuration (e.g., limo.yaml - LimoDataModule settings)
• model/: Model architecture (e.g., limo.yaml - network configuration)
• trainer/: PyTorch Lightning trainer settings
• logger/: Logging backends (e.g., wandb, csv)
• callbacks/: Training callbacks (e.g., checkpointing)
• experiment/: Complete experiment configs that override defaults

Experiment configs (limo/configs/experiment/):

• train_limo_debug.yaml: Quick debug run with 5 epochs, limited batches
• train_limo_on_D_aug.yaml: Full training on augmented dataset
• train_limo_side_cams.yaml: Training with side camera inputs

Override configs using the Hydra syntax:

# Use a specific experiment
uv run limo/src/train.py experiment=train_limo_on_D_aug

# Override specific parameters
uv run limo/src/train.py experiment=train_limo_debug data.batch_size=32 trainer.max_epochs=10

# Disable wandb logging
uv run limo/src/train.py experiment=train_limo_debug logger=null

By default, training uses Weights & Biases for logging.

Checkpoints and Weights

• Checkpoints: Saved to logs/train/runs/<timestamp>/checkpoints/ (PyTorch Lightning format)
• SafeTensors weights: Automatically converted and saved to logs/train/runs/<timestamp>/weights/ after each checkpoint

Dataset Builder

The dataset_builder package generates training samples from Grand Tour missions. It runs the MPPI geometric planner over pre-computed elevation maps to produce goal-conditioned paths, writing output in the same zarr format consumed by the training pipeline. Required topics are downloaded from HuggingFace automatically.

Quick start: build and train on 3 missions

# Build geo + tel paths for 3 missions - ETH outdoor, Jungfraujoch, Construction site
# (1 path/frame, skips first 1000 frames)
uv run dataset_builder/src/build_paths.py --config-name build_example dataset_type=geo
uv run dataset_builder/src/build_paths.py --config-name build_example dataset_type=tel

# Train on the result with aug (geo + tel), 5 epochs, CSV logger — no WandB needed
uv run limo/src/train.py experiment=train_limo_local_example

Output goes to data/dataset_builder/. experiment=train_limo_local_example selects dataset=limo_local_example (points at the three example missions, dataset_type=aug) and switches the logger to CSV.

Full build (all missions, as in the paper)

# D_geo: 10 paths per frame across all 48 missions
uv run dataset_builder/src/build_paths.py dataset_type=geo

# D_tel: teleoperation paths
uv run dataset_builder/src/build_paths.py dataset_type=tel

Then train on all missions:

uv run limo/src/train.py experiment=train_limo_local

Visualization

Pass viz=true to watch the planner and maps live while building:

uv run dataset_builder/src/build_paths.py --config-name build_example dataset_type=geo viz=true

The viewer shows elevation, traversability, and GDF maps in the bottom row, and camera images in the top row. Side camera images are included automatically if they were already downloaded (see below); if not, those panels are left blank without failing. The display updates every 5 frames (viz_every: 5); override with e.g. viz_every=1 on the command line.

Side cameras

By default only the front camera and odometry are downloaded. To also pull the left and right cameras — required for training the side-camera model variant — pass fetch_side_cams=true:

uv run dataset_builder/src/build_paths.py --config-name build_example dataset_type=geo fetch_side_cams=true
uv run dataset_builder/src/build_paths.py --config-name build_example dataset_type=tel fetch_side_cams=true

Side camera data is then available for experiment=train_limo_side_cams.

Note on data availability: The elevation maps (elevation_revision: refs/pr/9) and the pre-built LiMo path zarrs (HF_REVISION_LIMO = "refs/pr/6") live on open HuggingFace PRs against leggedrobotics/grand_tour_dataset that have not yet been merged to main. The dataset builder downloads from these PR branches automatically; they will be updated to point at main once the PRs are merged.

Dataset

LiMo's training data is based on Grand Tour dataset from HuggingFace: leggedrobotics/grand_tour_dataset We added LiMo's data to the same HuggingFace repo.

Note: The dataset is automatically pulled from HuggingFace. No manual download required!

Dataset Types

The Grand Tour dataset contains multiple mission recordings. Different sample types can be extracted:

• Teleoperation samples (tel): Expert demonstrations from human teleoperation
• Geometric samples (geo): Trajectories generated by the MPPI geometric planner
• Augmented samples (aug): Combined set of teleoperation + geometric samples

Select the dataset type using the get_dataset() method in dataset/src/limo_datset.py:

from dataset.src.limo_datset import get_dataset

# Load teleoperation samples only
dataset_tel = get_dataset(
    dataset_type="tel",           # or "geo", "aug"
    dataset_folder="data/dataset",
    missions_csv="missions_split.csv",
    with_side_cams=True
)

# Load augmented samples with side cameras
dataset_aug = get_dataset(
    dataset_type="aug",
    dataset_folder="data/dataset",
    missions_csv="missions_split.csv",
    with_side_cams=True
)

Missions and Train/Val/Test Split

The missions_split.csv file controls which missions are used and how they're split:

CSV Format:

Mission,Timestamp,Split
grandtour_mission_1,2024-01-15,train
grandtour_mission_2,2024-01-16,val
grandtour_mission_3,2024-01-17,test

Usage:

• Include only specific missions by adding rows to the CSV
• Control split ratios by adjusting the number of missions per split
• Quick start: use missions_split_example.csv (3 missions: train/val/test, used by train_limo_local)

Citation

If you use this work in your research, please cite:

@misc{inglin2026morescalablevisualnavigation,
      title={Less Is More: Scalable Visual Navigation from Limited Data},
      author={Yves Inglin and Jonas Frey and Changan Chen and Marco Hutter},
      year={2026},
      eprint={2601.17815},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2601.17815},
}