/crl_kino

Using RL based controller for steering function of RRT. And a network learns to perform motion planning

Primary LanguagePythonMIT LicenseMIT

This package is used to solve the kinodynamic problem using RRT.

Installation

Now it is still in development, to install it for development:

Notice that if you have installed tianshou, please remove it or create another virtual env to install this package. Because we use tianshou for reinforcemenr learning, and change it a bit for our use.

conda create -n crl_kino python=3.6 
conda activate crl_kino
git clone https://github.com/sldai/crl_kino.git
cd crl_kino
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
pip install -e .

Use

Test:

# kinodynamic RRT using dwa as steering function
python example/test.py --case rrt

# kinodynamic RRT using a trained policy as steering function
python example/test.py --case rl_rrt

Gif of kinodynamic rrt results will be saved in the current folder.

Train an agent:

python examples/rl_training.py

Results

RRT tree growing process path

The final course path