NTFields

NTFields: Neural Time Fields for Physics-Informed Robot Motion Planning
Ruiqi Ni, Ahmed H Qureshi

Paper | GitHub | arXiv | Published in ICLR 2023.

Introduction

This repository is the official implementation of "NTFields: Neural Time Fields for Physics-Informed Robot Motion Planning".

Installation

Clone the repository into your local machine:

git clone https://github.com/ruiqini/NTFields --recursive

Install requirements:

conda env create -f NTFields_env.yml
conda activate NTFields

Download datasets and pretrained models, exact and put datasets/ Experiments/ to the repository directory:

Datasets and pretrained model

The repository directory should look like this:

NTFields/
├── datasets/
│   ├── arm/    # 4DOF and 6DOF robot arm, table environment
│   ├── c3d/    # C3D environment
│   ├── gibson/ # Gibson environment
│   └── test/   # box and bunny environment
├── Experiments
│   ├── 4DOF/   # pretrained model for 4DOF arm
│   └── Gib/    # pretrained model for Gibson
•   •   •
•   •   •

Pre-processing

To prepare the Gibson data, run:

python dataprocessing/preprocess.py --config configs/gibson.txt

To prepare the arm data, run:

python dataprocessing/preprocess.py --config configs/arm.txt

Testing

To visualize our path in a Gibson environment, run:

python test/gib_plan.py

To visualize our path in the 4DOF arm environment, run:

python test/arm_plan.py

To sample random starts and goals in Gibson environments, run:

python test/sample_sg.py

To show our statistics result in Gibson environments, run:

python test/gib_stat.py

Training

To train our model in a Gibson environment, run:

python train/train_gib.py

To train our model in the 4DOF arm environment, run:

python train/train_arm.py

Docker

Build the docker image ntf-image by running:

cd docker && bash build-image.sh

Run the docker container as follows (let we define docker run --gpus all -v /path/to/NTFields-repo:/code-wkdir ntf-image as docker-ntf for brevity):

# preprocessing
docker-ntf dataprocessing/preprocess.py --config configs/gibson.txt
# training
docker-ntf train/train_gib.py

Citation

Please cite our paper if you find it useful in your research:

@inproceedings{
    ni2023ntfields,
    title={{NTF}ields: Neural Time Fields for Physics-Informed Robot Motion Planning},
    author={Ruiqi Ni and Ahmed H Qureshi},
    booktitle={International Conference on Learning Representations},
    year={2023},
    url={https://openreview.net/forum?id=ApF0dmi1_9K}
}

Acknowledgement

License

NTFields is released under the MIT License. See the LICENSE file for more details.

stalhabukhari/NTFields