MBPO PyTorch

A PyTorch reimplementation of MBPO (When to trust your model: model-based policy optimization)

Deprecated Warning

The owner of this repo has graduated and this repo is no longer maintained. Please refer to this new MBPO Pytorch re-implementation, which is a submodule of the Unstable Baselines project maintained by researchers from the same lab. This new MBPO re-implementation strictly follows the original TF implementation and has been tested on several MuJoCo tasks.

Dependency

Please refer to ./requirements.txt.

Usage

pip install -e .

# default hyperparams in ./configs/mbpo.yaml
# remember to CHANGE proj_dir to your actual directory 
python ./mbpo_pytorch/scripts/run_mbpo.py

# you can also overwrite hyperparams by passing args, e.g.
python ./mbpo_pytorch/scripts/run_mbpo.py --set seed=0 verbose=1 device="'cuda:0'" env.env_name='FixedHopper'

Credits

1. vitchyr/rlkit
2. JannerM/mbpo
3. WilsonWangTHU/mbbl