mujoco_metaworld_validation

Python environment for using Vintix

Building docker image

Run the following command to build the image:

docker build --platform linux/amd64 -t vintix-val-image:version1 .

Manual installation

For inference Vintix model numpy>=1.24.4, torch and flash-attn packages are required. However, ensure that flash-attn is compatible with the installed torch and CUDA version. For example, for CUDA 11.8, you can use the following command:

pip3 install torch --index-url https://download.pytorch.org/whl/cu118
pip3 install --no-cache-dir --no-build-isolation \
    ninja \
    setuptools \
    packaging \
    pydantic \
    wheel
pip3 install --no-cache-dir --no-build-isolation flash-attn

gymnasium>=0.28.0, mujoco<3 and imageio are required to create MuJoCo tasks.

To create Meta-World tasks using gym.make, you should install the Meta-World package:

pip3 install "metaworld @ git+https://github.com/qgallouedec/[email protected]_register"

The Meta-World requires the mujoco_py package, which depends on MuJoCo binaries: more information here.

wget https://mujoco.org/download/mujoco210-linux-x86_64.tar.gz
tar -xf mujoco210-linux-x86_64.tar.gz
mkdir $HOME/.mujoco
mv mujoco210 $HOME/.mujoco/

After installation some environment variables should be updated:

export MUJOCO_PY_MUJOCO_PATH="$HOME/.mujoco/mujoco210"
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco210/bin"
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/nvidia

Moreover, to compile this library, you must install cython<3. It’s also recommended to install the following Linux packages, as you may encounter compilation errors without them:

sudo apt-get update --fix-missing && sudo apt-get upgrade -y &&\
 sudo apt-get install -qy\
 python3-dev\
 build-essential\
 libblas-dev\
 libatlas-base-dev\
 libosmesa6-dev\
 libgl1-mesa-glx\
 libglfw3\
 patchelf\
 libgmp-dev\
 libgmpxx4ldbl\
 libglew-dev

For Ubuntu versions newer than 23.10, you need to replace libgl1-mesa-glx with libgl1 and libglx-mesa0 in the installation command above.

metaworld package should be imported before gymnasium to create Meta-World tasks using gymnasium.make:

import metaworld
import gymnasium as gym

env = gym.make('assembly-v2')

The versions of the packages the model was validated on are:

• CUDA 11.8
• Python 3.11.5
• Package versions

  torch==2.5.1+cu118
  flash-attn==2.7.0.post2
  numpy==1.24.1
  gymnasium==1.0.0
  mujoco==2.3.7

• Meta-World installed from repository mentioned above