roadsong / pi-prl

ICLR'22 Spotlight Programmatic Reinforcement Learning without Oracles

Programmatic Reinforcement Learning without Oracles

Wenjie Qiu, He Zhu

Rutgers University

ICLR 2022 @ Spotlight

[OpenReview] [Project Homepage]

Installation

We tested our code on Ubuntu 18.04 LTS x86_64 platform. Please make sure to read below instructions before using requirements.txt.

• Required Dependencies

  $ sudo apt install libopenmpi-dev libosmesa6-dev libgl1-mesa-glx libglfw3 libgl1-mesa-dev patchelf
  

• Install MuJoCo

  Download zip:

  $ wget https://roboti.us/download/mujoco200_linux.zip
  

  Create mujoco directory on your home folder:

  $ mkdir ~/.mujoco
  

  Unzip and move:

  $ unzip ./mujoco200_linux.zip 
  $ mv ./mucjoco200_linux ~/.mujoco/
  
  $ cd ~/.mujoco 
  $ mv ./mujoco200_linux mujoco200
  

  Download the key:

  $ wget https://roboti.us/file/mjkey.txt
  

  Check if the MuJoCo library and key are ready:

  $ ls ~/.mujoco
  mjkey.txt mujoco200
  

  Export to library path:

  $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco200/bin
  

  Upgrade pip (required):

  $ python3 -m pip --upgrade
  

  Notes: at this point, you may use requirements.txt to install the reamaining python packages.

  Install mujoco-py:

  $ pip3 install mujoco-py==2.0.2.13
  

  Now you should be able to import mujoco_py.

• Install Python Packages

  Tested on Python 3.6.9 (default version on Ubuntu 18.04).

  Install python packages:

  $ pip3 install numpy torch matplotlib z3-solver pillow tqdm tabulate joblib gym pathos==0.2.8 mpi4py==3.1.1
  

Architecture Search for Programmatic Policies (π-PRL)

Solving Ant Cross Maze (0), Ant Random Goal (1), HalfCheetah Hurdle (2) and Pusher (3) environments

$ python3 pi_PRL.py -e [environment number, default=0] -s [random seed, default=123] -d [directory to save, default=data]

Programmatic High-level Planning (π-HPRL)

Solving Ant Maze (0), Ant Push (1) and Ant Fall (2) environments.

$ python3 pi_HPRL.py -e [environment number, default=0] -s [random seed, default=123] -d [directory to save, default=data]

References

• OpenAI Spinningup: https://github.com/openai/spinningup
• mjrl: https://github.com/aravindr93/mjrl