Dreamer

This is a Pytorch implementation of paper: Dreamer to control: Learning Behaviors by Latent Imagination. Get similar results compared with original Tensorflow implementation. Tested on dm_control suite, with tasks cartpole-balance, cheetah-run, and ball-in-cup_catch. Testing on more tasks is ongoing.

Requirements

• Install dm_control and MuJoCo: following DeepMind Control Suite and MuJoCo
• Run conda env create -f environment.yml to install dependencies.

To train the model:

python3 dreamer.py --env [env-name]

More optional environments are listed in env.py . The code is only tested on dm_control suite. If you have any question, feel free to post issues.

Some training results:

• Robustness of action repeat

References:

[1] Dream to Control: Learning Behaviors by Latent Imagination

[2] Tensorflow implementation, with tensorflow1.0

[3] Tensorflow implementation, with tensorflow2.0

[4] Learning Latent Dynamics for Planning from Pixels

[5] PlaNet implementation from @Kaixhin