ScaleNav/docs/INSTALL.md at main · MachEmbodied/ScaleNav · GitHub
Skip to content

INSTALL.md

Latest commit

 

History

History
164 lines (123 loc) · 5.33 KB

INSTALL.md

File metadata and controls

164 lines (123 loc) · 5.33 KB
Raw

Installation

For navigation tests, complete Environment, Simulator installation, Scene converter and Checkpoints in order.

Hardware

Use Linux and an NVIDIA GPU with BF16 support. We recommend 24 GB GPU memory and 64 GB system RAM for inference and evaluation. See training requirements for training.

Environment

Create the Python 3.12 environment:

conda create -n scalenav --override-channels -c conda-forge python=3.12 pip ffmpeg git
source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate scalenav

git -c http.sslBackend=openssl clone https://github.com/MachEmbodied/ScaleNav.git
cd ScaleNav
python -m pip install --upgrade pip
python -m pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124
python -m pip install -e '.[inference,eval,training]'
awm doctor --require-cuda --data-tools

Install Miniforge if Conda is unavailable.

Simulator installation

Required for navigation evaluation and simulation-data rendering. After creating the environment, complete both steps below.

Build dependencies

Install the compiler and CUDA toolkit for Habitat-GS and DeepSpeed in the same scalenav environment:

conda activate scalenav
conda install --override-channels -c conda-forge \
  gcc_linux-64=12 gxx_linux-64=12 libegl-devel libgl-devel libjpeg-turbo zlib glm=0.9.9.8
conda install --override-channels -c nvidia/label/cuda-12.4.1 -c conda-forge cuda-toolkit=12.4.1
conda activate scalenav

python -m pip install cmake==3.31.10 ninja scikit-build-core pybind11 setuptools wheel

Build Habitat-GS

Run from the repository root. Habitat-GS provides the habitat_sim module.

(
  set -e
  conda activate scalenav
  export PYTHONNOUSERSITE=1
  export CUDA_HOME="${CONDA_PREFIX:?Activate the scalenav Conda environment}"
  export CUDACXX="$CUDA_HOME/bin/nvcc"
  export CUDAHOSTCXX="${CXX:?Install the build dependencies and reactivate scalenav}"
  export CMAKE_PREFIX_PATH="$CONDA_PREFIX${CMAKE_PREFIX_PATH:+:$CMAKE_PREFIX_PATH}"
  export CMAKE_BUILD_PARALLEL_LEVEL=2
  python -c 'import os, sys; assert sys.prefix == os.environ["CONDA_PREFIX"], "Python must come from the active Conda environment"; print(sys.executable)'
  python -m awm.cli doctor --require-cuda --data-tools
  "$CUDACXX" --version
  "$CUDAHOSTCXX" --version

  scalenav_git_slot=${GIT_CONFIG_COUNT:-0}
  export "GIT_CONFIG_KEY_${scalenav_git_slot}=http.sslBackend"
  export "GIT_CONFIG_VALUE_${scalenav_git_slot}=openssl"
  export GIT_CONFIG_COUNT=$((scalenav_git_slot + 1))

  if [ ! -d external/habitat-gs ]; then
    git clone https://github.com/zju3dv/habitat-gs.git external/habitat-gs
  fi
  git -C external/habitat-gs checkout cb01b2fdf850ce3cab28ad00461a68ac7719f23f
  git -C external/habitat-gs submodule update --init --recursive
  HABITAT_BUILD_GUI_VIEWERS=OFF HABITAT_WITH_CUDA=ON HABITAT_WITH_BULLET=OFF \
    python -m pip install ./external/habitat-gs --no-build-isolation \
    --config-settings=cmake.define.CMAKE_CUDA_ARCHITECTURES=native
  python -m awm.cli doctor --require-cuda --data-tools --simulator
)

Scene converter

Required for Habitat-GS-LN and simulation training data. Run from the repository root:

conda install --override-channels -c conda-forge nodejs=22
npm install --prefix external/splat-tools @playcanvas/splat-transform@3.10.1
export PATH="$PWD/external/splat-tools/node_modules/.bin:$PATH"

Checkpoints

Model Download Size
ScaleNav Hugging Face 10.02 GB
Wan VAE, text encoder and tokenizer Wan2.2-TI2V-5B 14.20 GB

Download both models and set the paths used by the commands below:

python scripts/download_checkpoint.py --output checkpoints/scalenav
awm check --checkpoint checkpoints/scalenav --sha256

python - <<'PYTHON'
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="Wan-AI/Wan2.2-TI2V-5B",
    revision="921dbaf3f1674a56f47e83fb80a34bac8a8f203e",
    local_dir="checkpoints/Wan2.2-TI2V-5B",
    allow_patterns=["Wan2.2_VAE.pth", "models_t5_umt5-xxl-enc-bf16.pth", "google/umt5-xxl/*"],
)
PYTHON

export CHECKPOINT=checkpoints/scalenav
export BASE_MODEL=checkpoints/Wan2.2-TI2V-5B
awm doctor --checkpoint "$CHECKPOINT" --base-model "$BASE_MODEL" --require-cuda

Continue with the navigation quick test.

Optional dependencies

DeepSpeed

For training with --deepspeed, complete the build dependencies, then install:

python -m pip install -e '.[distributed]'

DeepSpeed builds optimizer extensions on first use. In a new terminal, activate scalenav and set CUDA_HOME="$CONDA_PREFIX" before training.

Go2 robot

Install the official Unitree SDK2 Python package in the environment running the robot adapter. Policy serving and the dry-run adapter use the base environment above.

Development

Install the test and package-building tools:

python -m pip install -e '.[dev]'

From the repository root, run:

python -m pytest
python -m build
python scripts/check_release.py