This is a contemporary PyTorch implementation of the Dreamer v1 algorithm, originally developed by Danijar Hafner et al. The original paper is accessible here.

This implementation draws heavily from the following projects:

• The original Dreamer v1 implementation
• dreamer-pytorch by Julius Frost
• dreamer-pytorch by Yusuke Urakami

The purpose of this project is solely for academic use.

Additionally, this project serves as a supplementary project for our CSC413 final project at the University of Toronto, where I am collaborating with the following wonderful people:

• Charlotte
• Blair
• Ray

Feel free to reach me out at [email protected].

Reproduce Results

Task	Original Dreamer	Our Dreamer
Acrobot Swingup	~260	269
Cartpole Swingup	~760	771
Cheetah Run	~770	554
Hopper Hop	~240	130
Finger Turn Easy	~550	585
Finger Turn Hard	~500	520
Finger Spin	~550	119 (best 378)

Development Environment Setup

Option 1

conda create -n dreamerv1-pytorch -f environment.yml

Option 2

conda create -n dreamerv1-pytorch python=3.11 -y
conda activate dreamerv1-pytorch
conda install -c conda-forge cudatoolkit=11.8 mesalib -y
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip3 install -r requirements.txt

Run Experiments

Modify configurations in main.py and then

export MUJOCO_GL="osmesa"
python3 main.py