Skip to content

Latest commit

 

History

History
57 lines (45 loc) · 2.18 KB

README.md

File metadata and controls

57 lines (45 loc) · 2.18 KB
Raw

Dreaming in goal-conditioned environments

This code is building upon the model-based reinforcement learning agent Dreamer:

@inproceedings{Dreamer,
  author    = {Danijar Hafner and
               Timothy P. Lillicrap and
               Jimmy Ba and
               Mohammad Norouzi},
  title     = {Dream to Control: Learning Behaviors by Latent Imagination},
  booktitle = {8th International Conference on Learning Representations, {ICLR} 2020,
               Addis Ababa, Ethiopia, April 26-30, 2020},
  year      = {2020}
}

Specifically, it extends the implementation by first author Danijar Hafner to also work on robotic goal-conditioned OpenAI gym environments, such as FetchReach-v1. Read more about the goal-conditioned environment suite in this OpenAI blog post.

Instructions

Create the conda environment with all dependencies:

conda env create --file conda-env.yml
conda activate dreamer-env

This already installs the requirements in requirements.txt for you. Make sure you have MuJoCo set up on your machine beforehand (typically in /home/yourname/.mujoco/mujoco_200/). This is not done by conda for you! Besides the steps in the MoJoCo documentation, I also had to run the following commands inside of the conda env:

export LD_LIBRARY_PATH=$HOME/.mujoco/mujoco200/bin:$LD_LIBRARY_PATH
export MUJOCO_PY_MJPRO_PATH=$HOME/.mujoco/mujoco200/
export MUJOCO_PY_MJKEY_PATH=$HOME/.mujoco/mjkey.txt
sudo apt install libosmesa6-dev

Maybe libosmesa6-dev could be included inside conda-env.yml, but I was not able to find a suitable channel for it.

Train the agent using the specified logdir and robotics environment:

python3 dreamer.py --logdir ./logdir/fetch-reach-v1/dreamer/1 --task robotics_FetchReach-v1

Generate plots:

python3 plotting.py --indir ./logdir --outdir ./plots --xaxis step --yaxis test/return --bins 3e4

Start tensorboard with graphs and GIFs:

tensorboard --logdir ./logdir