MillionIntegrals / vel-miniworld

An example project using vel to train reinforcement learning agents on existing community gym environments. A work in progress repository.
MIT License
4 stars 1 forks source link

vel-miniworld

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:

How to run?

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.avi

Additional notes

For 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.

Some animations

Solving simple small gridworld environment:

Solving slightly more complex gridworld environment with sparse rewards:

Solving small 3D rendered world: