Implementation of the first version of the DCIL algorithm (paper) based on jax-based XPAG library.
📘 An improved version of the DCIL algorithm ((paper) is available here: https://github.com/AlexandreChenu/DCIL_XPAG
- Clone DCIL repo,
git clone https://github.com/AlexandreChenu/DCIL_I_XPAG.git- Create virtual environment dcil_env from environment.ylm,
cd DCIL_XPAG
conda env create --name dcil_env --file environment.yml- Clone + install XPAG (+ Jax),
git clone https://github.com/perrin-isir/xpag.git
cd xpag
git checkout 9ef7dd74b74fc71cee83c6a476adfebe4b977814
pip install -e .Check this Repo for more instructions.
- Clone + install maze or humanoid environments
git clone https://github.com/AlexandreChenu/gmaze_dcil.gitOR
git clone https://github.com/AlexandreChenu/gfetch_dcil.gitOR
git clone https://github.com/AlexandreChenu/ghumanoid_dcil.gitand
cd <env_directory>
pip install -e .
```-e .To run DCIL in the fetch environment, please clone Go-Explore repo:
git clone https://github.com/uber-research/go-explore.gitmove to the robustified directory:
cd go-explore/robustifiedFrom there, clone uber-research/atari-reset:
git clone https://github.com/uber-research/atari-reset.git atari_resetFinally, update the PYTHONPATH to include robustified:
export PYTHONPATH=$PYTHONPATH:path_to_goexplore/go-explore/robustifiedwith path_to_goexplore, the absolute path to the go-explore repo.
python test_DCIL_I_XPAG_dubins.py --demo_path ./demos/dubins_convert/1.demo --save_path </path/to/save/path>python test_DCIL_I_XPAG_fetch.py --demo_path ./demos/fetch_convert/6.demo --save_path </path/to/save/path>python test_DCIL_I_XPAG_humanoid.py --demo_path ./demos/humanoid_convert/1.demo --save_path </path/to/save/path> --eps_state 0.5 --value_clipping 1- trajs_it_- : training rollouts + skill-chaining evaluation
- value_skill_-it- : value for x-y position of skill starting state for different orientations
- transitions_- : sampled training transitions + segment between true desired goal and relabelled desired goal
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent