The MuJoCo Grasping Simulator framework is presented in the paper https://arxiv.org/pdf/2410.18835.

git clone https://github.com/boschresearch/mj-grasp-sim.git
cd mj-grasp-sim

conda create -n mj-grasp-sim python=3.11 -y
conda activate mj-grasp-sim

Replace /path/to/your/in, /path/to/your/out with your desired local directories.

mkdir -p $CONDA_PREFIX/etc/conda/activate.d
cat <<EOF > $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
export MGS_INPUT_DIR=/path/to/your/in
export MGS_OUTPUT_DIR=/path/to/your/out
EOF

mkdir -p /path/to/your/in
mkdir -p /path/to/your/out

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

Download the YCB dataset and place it under /asset/mj-objects/. Then download the GoogleScannedObjects (GSO) dataset and place it under /asset/mj-objects/. Next, run the gso_to_delete.py script on the GSO dataset to remove object sets that behave unrealistically during simulation. For container usage, we highly recommend excluding these asset files from the container image and mounting them into the container at deployment time to keep the image smaller and maintain flexibility.

mj-grasp-sim provides several modules accessible via Hydra for flexible configuration.

• Gripper Scan

  python -m mgs.cli.scan_gripper <config>

• Grasp Generation

  python -m mgs.cli.gen_grasps <config>

• Clutter Scene Generation

  python -m mgs.cli.gen_clutter_scene <config>

• Scene Rendering

  python -m mgs.cli.render_scene <config>

*Replace <config> with the desired Hydra configuration

• All modules support Hydra for configuration management, allowing you to customize settings via configuration files or command-line overrides.
• Ensure that the specified input and output directories are correctly set in the environment variables.

MuJoCo Grasping Simulator is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.

For a list of other open source components included in MuJoCo Grasp Simulator, see the file 3rd-party-licenses.txt.