RESCUE: Reasoning over Earth Scenes for Coordinated UAS-UGS Exploration

Submodules

This repo vendors lseg-minimal (dense CLIP / LSeg features) and map-anything (MapAnything 3D reconstruction) as Git submodules under lseg-minimal/ and map-anything/.

Install

1. Pull submodules (from the RESCUE repo root):

   • New clone: git clone --recurse-submodules <RESCUE-repo-url>
   • Existing clone: git submodule update --init --recursive

2. Conda environment

   conda create -n rescue python=3.12 -y
   conda activate rescue

3. Dependencies

   pip install torch==2.10.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
   pip install git+https://github.com/openai/CLIP.git

4. LSeg:

   cd lseg-minimal
   python setup.py build develop
   cd ..

   LSeg checkpoint (Google Drive):

   pip install gdown
   mkdir -p generated
   gdown 1jGDqefX057MFvYIZN25khAhdS2AGFjjA -O generated/lseg_minimal_e200.ckpt

5. MapAnything

   cd map-anything
   ## use pip install -e ".[all]" for all dependencies to run vggt etc
   pip install -e . 
   cd ..

6. Download Renders

   The dataset renders can be downloaded using

   pip install gdown
   mkdir -p generated/renders
   gdown --folder https://drive.google.com/drive/folders/1zqisevgIJ_X6RczFZE54lGKOAFn1UheX -O generated/renders

   This will download all files in the renders folder to generated/renders/.

7. rescue as a library (from the RESCUE repo root, after the steps above):

   pip install -e .