Codebase for NeurIPS 2022: "Pre-Trained Image Encoder for Generalizable Visual Reinforcement Learning"
The MuJoCo license and instruction can be found at https://github.com/deepmind/mujoco;The Mujoco Version should be 2.1.1
for training.
The DeepMind Control license and instruction can be found at https://github.com/deepmind/dm_control
For training, the dependencies are based on DrQ-v2. You can install them with the following commands:
conda env create -f conda_env.yml
Detailed installation instructions can be found at: https://github.com/facebookresearch/drqv2
For generalization testing, we use the DMControl Gneralization Benchmark. You can run the commands as follows:
cd dmcontrol-generalization-benchmark/
conda env create -f setup/dmcgb.yml
conda activate dmcgb
sh setup/install_envs.sh
The place365 dataset can be downloaded by running:
wget http://data.csail.mit.edu/places/places365/places365standard_easyformat.tar
After downloading and extracting the data, add your dataset directory to the config.cfg
.
Detailed installation instructions can be found at: https://github.com/nicklashansen/dmcontrol-generalization-benchmark
pieg
conda environment is served for training, so you should activate this conda env at first:
conda activate pieg
bash pieg_train.sh task=walker_walk seed=1
cd
to the exp_local
file and move the trained model to the test file:
mv snapshot.pt ~/PIE-G/dmcontrol-generalization-benchmark/logs/walker_walk/pieg/1
cd ~/PIEG/dmcontrol-generalization-benchmark/
conda activate dmcgb
bash script/eval/pieg.sh 1 video_hard walker walk
The majority of PIE-G 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.