An example project using vel to train reinforcement learning agents on existing community gym environments. A work-in-progress repository.
Supported environments:
Examples confirmed to be working:
examples-configs/ppo/ppo_minigrid_empty_8x8.yamlexamples-configs/ppo/ppo_minigrid_doorkey_6x6.yaml(doesn't converge every time)examples-configs/ppo/ppo_miniworld_hallway.yaml
git clone git@github.com:MillionIntegrals/vel-miniworld.git
cd vel-miniworld
# Optionally, if you don't want to store metrics in the db and visualize in VisDom
mv .velproject.dummy.yaml .velproject.yaml
pipenv install
pipenv shell
vel examples-configs/ppo/ppo_minigrid_empty_8x8.yaml train
vel examples-configs/ppo/ppo_minigrid_empty_8x8.yaml record
# Optionally, play a video of agent solving a rather simple environment
mplayer output/videos/ppo_minigrid_empty_8x8/0/ppo_minigrid_empty_8x8_vid_0010.aviFor the textures to load properly for the 3D rendered miniworld environment, it needs to be installed
from a git repository, by running pip install -e . in the top-level directory of the checkout.
Let me know if you have any other problems running the environments.
Solving simple small gridworld environment:
Solving slightly more complex gridworld environment with sparse rewards:
Solving small 3D rendered world:
