This repo implements the Dreamer algorithm from Dream to Control: Learning Behaviors By latent Imagination based on the PlaNet-Pytorch. It has been confirmed working on the DeepMind Control Suite/MuJoCo environment. Hyperparameters have been taken from the paper.
To install all dependencies with Anaconda run using the following commands.
conda env create -f conda_env.yml
source activate dreamer
python main.py --algo dreamer --env walker-walk --action-repeat 2 --id name-of-experiementFor best performance with DeepMind Control Suite, try setting environment variable MUJOCO_GL=egl (see instructions and details here).
Use Tensorboard to monitor the training.
tensorboard --logdir results
The performances are compared with the other SoTA algorithms as follows (Note! Tested once using seed 0.)
NOTE! All the steps below are environment steps. (All the results below are trained 1000 episodes. Each espisode length is 1000 steps. Actual number of sampled data are steps/action_repeat. )
Pretrained models can be found in the releases.