CorrGen

CorrGen — A Differentiable Collision-Free Corridor Generator.

Camera	Point cloud with corridors

For the implementation details, please check the paper and/or watch the video.

If you use this framework please cite our work:

@misc{arrizabalaga2024differentiablecollisionfreeparametriccorridors,
      title={Differentiable Collision-Free Parametric Corridors}, 
      author={Jon Arrizabalaga and Zachary Manchester and Markus Ryll},
      year={2024},
      eprint={2407.12283},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2407.12283}, 
}

Quickstart

Install dependencies with

sudo apt-get install libcdd-dev

Create a python environment with python 3.9. For example, with conda:

conda create --name corrgen python=3.9
conda activate corrgen
pip install -r requirements.txt

Update the ~/.bashrc with

export CORRGEN_PATH=/path_to_pfdq
export PYTHONPATH=$PYTHONPATH:/$CORRGEN_PATH

Usage

KITTI dataset

To run a real-world example from the KITTI dataset (Figs 5 and 6 in the paper), run this command:

python examples/kitti.py --case p  --lp --n_corrgen 6 --n_decomp 6

The options are the following ones:

--case: p (pink corridor) or g (green corridor).
--lp: Runs the approximated LP instead of the original SDP
--n_corrgen: Integer indicating the polynomial degree of the polynomials in corrgen.
--n_decomp: Sets the number of polygons for convex decomposition (and runs it)

Toy example

To run a toy example (Fig. 4 in the paper), run this command:

python examples/toy_example.py --lp --n_corrgen 6

The options are the same as for the KITTI example (except for --case).

2D cross section comparison

To run the comparison of using different cross section parameterizations (Fig.3 in the paper), run this command:

python examples/cross_section.py

Notice that every time you run the command, the point cloud in the cross section varies. This is a great standalone script, great for conceptual prototyping.

Related repositories

For a discrete representation of the collision-free space via convex decomposition, check out pydecomp!