Dataset Setup

1. Download the datasets from here

2. Extract them maintaining the following structure:

   /path/to/your/data/
   └── <bin_size>_bins/
       ├── finger-spin-expert
       ├── fish-swim-expert
       └── ...
   

3. Set the environment variable DMC_DISCRETE_DATA_DIR:

   export DMC_DISCRETE_DATA_DIR=/path/to/your/data

   For a temporary setting, run this in your terminal. To make it permanent, add the export line to your shell's configuration file:

   For Bash: ~/.bashrc or ~/.bash_profile
   For Zsh: ~/.zshrc
   

   Or pass the data_dir parameter directly to load_dataset().

4. Git clone this repo and and use pip intsall -e . from the root of the repo

Dataset Usage

The datasets can be loaded as follows:

from dmc_datasets.environment_utils import load_dataset

dataset = load_dataset(task_name='finger', task='spin', bin_size=3, level='medium')

dataset will be a list of transitions.

Currently we only support bin size of 3 for cheetah-run, finger-spin, fish-swim, quaduped-walk, humanoid-stand, dog-trot and bin sizes {10, 30, 50, 75, 100} for dog-trot. Each dataset has a medium, expert, medium-expert and random-medium-expert level.

Environment wrappers

1. We also provide wrappers for the DMC dataset that can be used to return gym environments with either factorised or atomic discrete action spaces:

   from dmc_datasets.environment_utils import make_env
   
   env = make_env(task_name='finger', task='spin', bin_size=3, factorised=True)

2. This environment class can also be used to load a corresponding dataset for the environment instance:

   from dmc_datasets.environment_utils import make_env
   
   env = make_env(task_name='finger', task='spin', bin_size=3, factorised=True)
   dataset = env.load_dataset(level='medium', return_type='raw')  # return_type supports 'raw', 'dict', 'replay_buffer'