Hierarchical Diffusion Policy

Hierarchical Diffusion Policy: Manipulation Trajectory Generation Via Contact Guidance

Dexin Wang, Chunsheng Liu, Faliang Chang, Yichen Xu

This paper has been accepted by IEEE Transactions on Robotics (TRO).

[Paper] [Data] [Video]

🛠️ Installation

🖥️ Simulation (Same as Diffusion Policy)

To reproduce our simulation benchmark results, install conda environment on a Linux machine with Nvidia GPU. On Ubuntu 20.04 you need to install the following apt packages for mujoco:

$ sudo apt install -y libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf

We recommend Mambaforge instead of the standard anaconda distribution for faster installation:

$ mamba env create -f conda_environment.yaml

but you can use conda as well:

$ conda env create -f conda_environment.yaml

🦾 Real Robot

Data and code are being collated.

Software:

• Ubuntu 20.04.3 (tested)
• Mujoco dependencies: sudo apt install libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf
• Conda environment mamba env create -f conda_environment_real.yaml

🖥️ Reproducing Simulation Benchmark Results

Download Training Data

Under the repo root, download data.tar.xz and decompress:

[hierarchical_diffusion_policy]$ pip install gdown
[hierarchical_diffusion_policy]$ gdown https://drive.google.com/uc?id=1D9OkLOwDUjRjjRD6fKhF0p-tXz4KfXtW
[hierarchical_diffusion_policy]$ tar -xvJf data.tar.xz

Running for a single seed

Activate conda environment and login to wandb (if you haven't already).

[hierarchical_diffusion_policy]$ conda activate robodiff
(robodiff)[hierarchical_diffusion_policy]$ wandb login

Launch training with seed 42 on GPU 0.

(robodiff)[hierarchical_diffusion_policy]$ python train_hdp.py --config-dir=. --config-name=hdp_tilt.yaml training.seed=42 training.device=cuda:0 hydra.run.dir='data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${logging.name}'

This will create a directory in format data/outputs/yyyy.mm.dd/hh.mm.ss_<method_name>_<task_name> where configs, logs and checkpoints are written to. Guider, Critic and Actor will be trained in sequence. Actor will be evaluated every 50 epochs with the success rate logged as test/mean_score on wandb, as well as videos for some rollouts.

🆕 Evaluate Pre-trained Checkpoints

Download checkpoints of Guider and Actor from the published checkpoint, such as Actor ckpt and Guider ckpt.

Run the evaluation script:

(robodiff)[hierarchical_diffusion_policy]$ python eval_hdp.py --config-dir=. --config-name=hdp_tilt.yaml guider_path='guider_ckpt_path' actor_path='actor_ckpt_path' training.device=cuda:0 hydra.run.dir='data/eval'

Dataset

Robomimic

Run the following code to augment the Robomimic dataset published by Diffusion Policy with object point clouds. The augmented dataset with object point clouds will be saved in the same directory. For example, running the code below will generate a new dataset named low_dim_abs_pcd.hdf5 in the path data/robomimic/datasets/square/mh/.

(robodiff)[hierarchical_diffusion_policy]$ python dataset/add_pcd_to_robomimic_dataset.py 'data/robomimic/datasets/square/mh/low_dim_abs.hdf5'

Cite

@ARTICLE{10912754,
  author={Wang, Dexin and Liu, Chunsheng and Chang, Faliang and Xu, Yichen},
  journal={IEEE Transactions on Robotics}, 
  title={Hierarchical Diffusion Policy: Manipulation Trajectory Generation Via Contact Guidance}, 
  year={2025},
  volume={},
  number={},
  pages={1-20},
  doi={10.1109/TRO.2025.3547272}}