This guide matches the current BridgePolicy codebase and folder structure.

The setup below is a practical reference for Linux machines with NVIDIA GPUs. Please make sure your PyTorch build matches your local CUDA version.

If you have already cloned the repository, just cd into it. Otherwise:

git clone <your-bridgepolicy-repo-url>
cd bridge_policy

Please follow the steps in order. In particular, keep the Gym version consistent with the local copy in third_party/gym-0.21.0.

1. Create a Python environment

conda remove -n bridgepolicy --all
conda create -n bridgepolicy python=3.8
conda activate bridgepolicy

2. Install PyTorch

# Example: CUDA 12.1
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# Otherwise, install the torch version that matches your CUDA runtime.

3. Install BridgePolicy

cd BridgePolicy
pip install -e .
cd ..

This installs the main package:

bridge_policy_3d

4. Install MuJoCo in ~/.mujoco

cd ~/.mujoco
wget https://github.com/deepmind/mujoco/releases/download/2.1.0/mujoco210-linux-x86_64.tar.gz -O mujoco210.tar.gz --no-check-certificate
tar -xvzf mujoco210.tar.gz

Add the following to your shell config (for example ~/.bashrc), then source ~/.bashrc and open a new terminal:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${HOME}/.mujoco/mujoco210/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/nvidia
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
export MUJOCO_GL=egl

Then install mujoco-py from third_party:

cd third_party/mujoco-py-2.1.2.14
pip install -e .
cd ../..

5. Install simulation environments

pip install setuptools==59.5.0 Cython==0.29.35 patchelf==0.17.2.0

cd third_party
cd dexart-release && pip install -e . && cd ..
cd gym-0.21.0 && pip install -e . && cd ..
cd Metaworld && pip install -e . && cd ..
cd rrl-dependencies && pip install -e mj_envs/. && pip install -e mjrl/. && cd ..
cd ..

Optional assets:

• If you plan to run DexArt tasks, download the DexArt assets and place them under third_party/dexart-release/assets.
• If you plan to run Adroit experiments that depend on expert checkpoints, place the required checkpoints under the corresponding local third-party folder before training.

6. Install simplified PyTorch3D

cd third_party/pytorch3d_simplified
pip install -e .
cd ../..

7. Install additional Python packages

pip install zarr==2.12.0 wandb ipdb gpustat dm_control omegaconf hydra-core==1.2.0 dill==0.3.5.1 einops==0.4.1 diffusers==0.11.1 numba==0.56.4 moviepy imageio av matplotlib termcolor

If you use the UniDB-related components, also install:

pip install torchsummaryX lmdb lpips numpy opencv-python Pillow PyYAML scipy tensorboardX timm tqdm gradio tensorboard ema_pytorch IPython pytorch_fid

Note: the codebase imports einops, so use the einops package name rather than einop.

8. Install the point cloud visualizer (optional)

pip install kaleido plotly
cd visualizer
pip install -e .
cd ..

9. Before training

BridgePolicy task configs intentionally leave dataset paths blank for open-source release. Before launching training, set the dataset path in one of the following ways:

• Edit the relevant file under BridgePolicy/bridge_policy_3d/config/task/ and fill in zarr_path
• Or override it from the command line with Hydra

Example:

cd BridgePolicy
python train.py --config-name=<your_config>.yaml task.dataset.zarr_path=/path/to/your_dataset.zarr

Training scripts in scripts/ write outputs under data/outputs by default. You can override the root output directory with:

export BRIDGE_POLICY_RUN_ROOT=/path/to/your/output_root

After that, you can use the provided shell scripts in scripts/ or launch training directly with Hydra.

If you find this repository useful in your research, please consider citing our paper:

@article{liu2025sample,
  title={Sample from What You See: Visuomotor Policy Learning via Diffusion Bridge with Observation-Embedded Stochastic Differential Equation},
  author={Liu, Zhaoyang and Pan, Mokai and Wang, Zhongyi and Zhu, Kaizhen and Lu, Haotao and Zhang, Haipeng and Wang, Jingya and Shi, Ye},
  journal={arXiv preprint arXiv:2512.07212},
  year={2025}
}