This repository contains the code for the Stanford NAV Lab's official solution for the 2025 Lunar Autonomy Challenge.

Structure

Our main agent is located under agents/nav_agent.py. This class is the entry point for the autonomy stack and defines the run_step method which is called every step of the simulator.

Setup

Clone this repo inside the unzipped LunarAutonomyChallenge folder provided by the organizers which contains the simulator:

  LunarAutonomyChallenge
    ...
    lunar_autonomy_challenge
    ...

Create an outputs/ folder to store generated data, and a data/ folder to store other data (heightmaps, etc.).

Environment

1. Create conda env

conda create -n lac python=3.10
conda activate lac

2. Setup LAC simulator

• Download simulator folder from LunarAutonomyChallenge.zip
• Unzip it into ~/ or desired location
• cd ~/LunarAutonomyChallenge
• pip3 install --force-reinstall -r requirements.txt

3. Install pytorch into lac environment: pip3 install torch torchvision torchaudio
4. Clone this repo into the LunarAutonomyChallenge folder. Inside ~/LunarAutonomyChallenge/lunar_autonomy_challenge:

pip install -r requirements.txt
pip install -e .

5. Install LightGlue. In ~/opt:

git clone https://github.com/cvg/LightGlue.git && cd LightGlue
python -m pip install -e .

6. pip install -U segmentation-models-pytorch

Conventions

Transformations

We use the GTSAM convention, where a_T_b denotes the transformation from frame b to frame a.

• Also equivalent to the pose of frame b in frame a.
• a_T_b * b_T_c = a_T_c
• a_T_b * b_P = a_P (where b_P is points P in frame b)