RL Algorithms for Visual Continuous Control

Pytorch implementations of

• SAC: Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor
• SAC+AE: Improving Sample Efficiency in Model-Free Reinforcement Learning from Images
• CURL: CURL: Contrastive Unsupervised Representations for Reinforcement Learning
• RAD: Reinforcement Learning with Augmented Data
• DrQ: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels
• ATC: Decoupling Representation Learning from Reinforcement Learning

Instructions

Install MuJoCo 2.1 if it is not already the case (please refer to https://github.com/deepmind/dm_control.

Create conda environment

conda create -n visualrl -f conda_env.yml

Run the code

conda activate visualrl
python src/train.py  --agent drq  --domain_name cheetah  --task_name run

Results

Here's the benchmark results on PlaNet Benchmark tasks. All results are averaged over 5 different seeds.

Citation

Please use the bibtex below if you want to cite this repository in your publications:

@misc{rlcodebase,
  author = {Jinwei Xing},
  title = {Pytorch Implementations of Reinforcement Learning Algorithms for Visual Continuous Control},
  year = {2022},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/KarlXing/RL-Visual-Continuous-Control}},
}