Neural Configuration-Space Barriers for Manipulation Planning and Control [Paper]

This repository contains the implementation for the paper "Neural Configuration-Space Barriers for Manipulation Planning and Control".

If you find our work useful, please consider citing our paper:

@article{long2025neural_cdf,
  title={Neural Configuration-Space Barriers for Manipulation Planning and Control},
  author={Long, Kehan and Lee, Ki Myung Brian and Raicevic, Nikola and Attasseri, Niyas and Leok, Melvin and Atanasov, Nikolay},
  journal={arXiv preprint arXiv:2503.04929},
  year={2025}
}

🚀 Quick Start

Clone the repository:

git clone https://github.com/KehanLong/cdf_bubble_control.git
cd cdf_bubble_control

📦 Dependencies

This code has been tested on Ubuntu 22.04 LTS. You can run it using either Docker (recommended) or conda environment.

🐳 Using Docker (Recommended)

1. Build the Docker image:

docker build -t bubble_cdf_planner:latest .

2. Make the run script executable:

chmod +x run_docker.sh

3. Run the container:

./run_docker.sh

This will start an interactive shell in the container. You can then run the examples as described below.

🐍 Using Conda (Alternative)

If you prefer using conda, you can set up the environment:

conda env create -f environment.yml
conda activate cdf_bubble_planning_control

If using conda environment, you will additionally need to install OMPL dependencies: https://ompl.kavrakilab.org/installation.html

2D 2-link Planar Robot

Neural CDF Training

Default training dataset is saved in 2Dexamples/cdf_training/data/. To train the neural CDF:

python 2Dexamples/cdf_training/cdf_train.py

The default trained model is saved in 2Dexamples/trained_models/.

Bubble-CDF Planning

To run the bubble-CDF planning, run the following command:

python 2Dexamples/main_planning.py

For a gif illustration of the bubble-CDF planner, run:

python 2Dexamples/main_planning_gif.py

Baseline Comparison

To compare with baselines using OMPL's planners, run the following command (Uncomment the appropriate line under main to run desired planner/controller):

python 2Dexamples/planning_benchmark.py

DRO-CBF Control

To run the DRO-CBF control, run the following command:

python 2Dexamples/main_control.py

MPPI Baseline

To run the MPPI (Model Predictive Path Integral) baseline for control:

python 2Dexamples/main_mppi.py

PyBullet (xArm6)

Neural SDF/CDF Training

The default trained SDF/CDF models are saved in xarm_pybullet/trained_models/. Pre-trained models are provided, so training is optional. If you want to retrain the models:

• SDF Training: python xarm_pybullet/sdf_training/train.py
• CDF Training: python xarm_pybullet/cdf_training/train_online_batch.py

Bubble-CDF Planning

To run the bubble-CDF planning in PyBullet:

# Default settings
python xarm_pybullet/xarm_planning.py

# Custom settings
python3 xarm_pybullet/xarm_planning.py --goal [0.8,0.1,0.68] --planner mppi --dynamic_obstacles True  --seed 42 --gui False --early_termination True

# Available planners: bubble, bubble_connect, sdf_rrt, cdf_rrt, lazy_rrt, rrt_connect, mppi
# early termination: Stop after first valid path or explore all goal configurations

DRO-CBF Control

To run the DRO-CBF control:

# Default settings (with dynamic obstacles)
python xarm_pybullet/xarm_control.py

# Custom settings
python xarm_pybullet/xarm_control.py --goal [0.7,0.1,0.6] --planner bubble --controller clf_dro_cbf --dynamic True --gui True --early_termination True

# Available options:
# - planners: bubble, sdf_rrt, cdf_rrt, rrt_connect, lazy_rrt ...
# - dynamic: Whether to use dynamic obstacles
# - controllers: pd, clf_cbf, clf_dro_cbf