D4RL: Datasets for Deep Data-Driven Reinforcement Learning

D4RL is an open-source benchmark for offline reinforcement learning. It provides standardized environments and datasets for training and benchmarking algorithms. A supplementary whitepaper and website are also available.

Setup

D4RL can be installed by cloning the repository as follows:

git clone https://github.com/rail-berkeley/d4rl.git
cd d4rl
pip install -e .

Or, alternatively:

pip install git+https://github.com/rail-berkeley/d4rl@master#egg=d4rl

The control environments require MuJoCo as a dependency. You may need to obtain a license and follow the setup instructions for mujoco_py. This mostly involves copying the key to your MuJoCo installation folder.

The Flow and CARLA tasks also require additional installation steps:

• Instructions for installing CARLA can be found here
• Instructions for installing Flow can be found here. Make sure to install using the SUMO simulator, and add the flow repository to your PYTHONPATH once finished.

Using d4rl

d4rl uses the OpenAI Gym API. Tasks are created via the gym.make function. A full list of all tasks is available here.

Each task is associated with a fixed offline dataset, which can be obtained with the env.get_dataset() method. This method returns a dictionary with:

• observations: An N by observation dimensional array of observations.
• actions: An N by action dimensional array of actions.
• rewards: An N dimensional array of rewards.
• terminals: An N dimensional array of episode termination flags. This is true when episodes end due to termination conditions such as falling over.
• timeouts: An N dimensional array of termination flags. This is true when episodes end due to reaching the maximum episode length.
• infos: Contains optional task-specific debugging information.

You can also load data using d4rl.qlearning_dataset(env), which formats the data for use by typical Q-learning algorithms by adding a next_observations key.

import gym
import d4rl # Import required to register environments

# Create the environment
env = gym.make('maze2d-umaze-v1')

# d4rl abides by the OpenAI gym interface
env.reset()
env.step(env.action_space.sample())

# Each task is associated with a dataset
# dataset contains observations, actions, rewards, terminals, and infos
dataset = env.get_dataset()
print(dataset['observations']) # An N x dim_observation Numpy array of observations

# Alternatively, use d4rl.qlearning_dataset which
# also adds next_observations.
dataset = d4rl.qlearning_dataset(env)

Datasets are automatically downloaded to the ~/.d4rl/datasets directory when get_dataset() is called. If you would like to change the location of this directory, you can set the $D4RL_DATASET_DIR environment variable to the directory of your choosing, or pass in the dataset filepath directly into the get_dataset method.

Normalizing Scores

You can use the env.get_normalized_score(returns) function to compute a normalized score for an episode, where returns is the undiscounted total sum of rewards accumulated during an episode.

The individual min and max reference scores are stored in d4rl/infos.py for reference.

Algorithm Implementations

We have aggregated implementations of various offline RL algorithms in a separate repository.

Off-Policy Evaluations

D4RL currently has limited support for off-policy evaluation methods, on a select few locomotion tasks. We provide trained reference policies and a set of performance metrics. Additional details can be found in the wiki.

Recent Updates

2-12-2020

• Added new Gym-MuJoCo datasets (labeled v2) which fixed Hopper's performance and the qpos/qvel fields.
• Added additional wiki documentation on generating datasets.

Acknowledgements

D4RL builds on top of several excellent domains and environments built by various researchers. We would like to thank the authors of:

• hand_dapg
• gym-minigrid
• carla
• flow
• adept_envs

Citation

Please use the following bibtex for citations:

@misc{fu2020d4rl,
    title={D4RL: Datasets for Deep Data-Driven Reinforcement Learning},
    author={Justin Fu and Aviral Kumar and Ofir Nachum and George Tucker and Sergey Levine},
    year={2020},
    eprint={2004.07219},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Licenses

Unless otherwise noted, all datasets are licensed under the Creative Commons Attribution 4.0 License (CC BY), and code is licensed under the Apache 2.0 License.