DrQ-v2: Improved Data-Augmented RL Agent

Method

DrQ-v2 is a model-free off-policy algorithm for image-based continuous control. DrQ-v2 builds on DrQ, an actor-critic approach that uses data augmentation to learn directly from pixels. We introduce several improvements including:

• Switch the base RL learner from SAC to DDPG.
• Incorporate n-step returns to estimate TD error.
• Introduce a decaying schedule for exploration noise.
• Make implementation 3.5 times faster.
• Find better hyper-parameters.

These changes allow us to significantly improve sample efficiency and wall-clock training time on a set of challening tasks from the DeepMind Control Suite compared to prior methods. Furthermore, DrQ-v2 is able to solve complex humanoid locomotion tasks directly from pixel observations, previously unattained by model-free RL.

Citation

If you use this repo in your research, please consider citing the paper as follows:

@article{yarats2021drqv2,
  title={Mastering Visual Continuous Control: Improved Data-Augmented Reinforcement Learning},
  author={Denis Yarats and Rob Fergus and Alessandro Lazaric and Lerrel Pinto},
  journal={arXiv preprint arXiv:},
  year={2021}
}

Instructions

Install dependencies:

conda env create -f conda_env.yml
conda activate drqv2

Train the agent:

python train.py task=quadruped_walk

Monitor results:

tensorboard --logdir exp_local

License

The majority of DrQ-v2 is licensed under the MIT license, however portions of the project are available under separate license terms: DeepMind is licensed under the Apache 2.0 license.