Motion Planning Diffusion: Learning and Adapting Robot Motion Planning with Diffusion Models

This repository implements Motion Planning Diffusion (MPD) - a method for learning and planning robot motions with diffusion models.

An older version of this project is deprecated, but still available at https://github.com/jacarvalho/mpd-public.

Please contact me if you have any questions -- joao@robot-learning.de

---

Installation

Pre-requisites:

• Ubuntu 22.04 (maybe works with newer versions)
• miniconda

Clone this repository

mkdir -p ~/Projects/MotionPlanningDiffusion/
cd ~/Projects/MotionPlanningDiffusion/
git clone --recurse-submodules git@github.com:joaoamcarvalho/mpd-splines-public.git mpd-splines-public
cd mpd-splines-public

Download IsaacGym Preview 4 and extract it under deps/isaacgym

mv ~/Downloads/IsaacGym_Preview_4_Package.tar.gz ~/Projects/MotionPlanningDiffusion/mpd-splines-public/deps/
cd ~/Projects/MotionPlanningDiffusion/mpd-splines-public/deps
tar -xvf IsaacGym_Preview_4_Package.tar.gz

Run the bash setup script to install everything (it can take a while).

bash setup.sh

Make sure to set environment variables and activate the conda environment before running any scripts.

source set_env_variables.sh
conda activate mpd-splines-public

---

Download the datasets and pre-trained models

Download https://drive.google.com/file/d/1KG5ejn0g0KkDuUK6tPUqfmRYCNoKzK4K/view?usp=drive_link

tar -xvf data_public.tar.gz
ln -s data_public/data_trajectories data_trajectories
ln -s data_public/data_trained_models data_trained_models

---

Inference with pre-trained models

The configuration files under scripts/inference/cfgs contain the hyperparameters for inference.
Inside the file scripts/inference/inference.py you can change the cfg_inference_path parameter to try models trained for different environments.

cd scripts/inference
python inference.py

---

Training the prior models (from scratch)

Data generation

Generating data takes a long time, so we recommend downloading the dataset. But if anyway you want to generate your own data, you can do it with the scripts in the scripts/generate_data folder.

Go to the scripts/generate_data folder.

The base script is

python generate_trajectories.py

To generate multiple datasets in parallel, adapt the launch_generate_trajectories.py script.

python launch_generate_trajectories.py

After generating the data, run the post-processing file to combine all data into a hdf5 file. Then you can double the dataset by flipping the trajectory paths.

python post_process_trajectories.py --help
python flip_solution_paths.py  (change the PATH_TO_DATASETS variable)

To visualize the generated data, use the visualize_trajectories.py script.

python visualize_trajectories.py

---

Training the models

The training scripts are in the scripts/train folder.

The base script is

cd scripts/train
python train.py

To train multiple models in parallel, use the launch_train_* files.

---

Citation

If you use our work or code base, please cite our articles:

@article{carvalho2025motion,
  title={Motion planning diffusion: Learning and adapting robot motion planning with diffusion models},
  author={Carvalho, Jo{\~a}o and Le, An T and Kicki, Piotr and Koert, Dorothea and Peters, Jan},
  journal={IEEE Transactions on Robotics},
  year={2025},
  publisher={IEEE}
}

@inproceedings{carvalho2023motion,
  title={Motion planning diffusion: Learning and planning of robot motions with diffusion models},
  author={Carvalho, Jo{\~a}o and Le, An T and Baierl, Mark and Koert, Dorothea and Peters, Jan},
  booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2023}
}

---

Credits

Parts of this work and software were taken and/or inspired from:

• https://github.com/jannerm/diffuser