Dreamer implementation in PyTorch

Dreamer-torch1.8.2

Updated the code to catch up with the latest pytorch version.

Previous version for pytorch 1.4 is left at dreamer branch.

Not conpatible with open-source Mujoco yet. Please download and use mujoco-200 and free license key from here

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.

Installation

To install all dependencies with Anaconda run using the following commands.

conda env create -f conda_env.yml

source activate dreamer

Training (e.g. DMC walker-walk)

python main.py --algo dreamer --env walker-walk --action-repeat 2 --id name-of-experiement

For 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

Results

The performances are compared with the other SoTA algorithms as follows (Note! Tested once using seed 0.)

• State-SAC
• PlaNet-PyTorch
• SAC-AE
• SLAC
• CURL
• Dreamer (tensorflow2 implementation)

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.

Links

• Dream to Control: Learning Behaviors By latent Imagination
• google-research/dreamer