SiMPL : Skill-based Meta-Policy Learning

[Project Page] [Paper]

Official implementation of the paper "Skill-based Meta-Reinforcement Learning", ICLR 2022.

Requirements

• Python 3.6+
• Mujoco
• D4RL
• pip packages listed in requirements.txt

Getting Started

1. Install simpl package from code

   cd /project/home
   mkdir lib && cd lib
   git clone https://github.com/namsan96/SiMPL.git
   cd SiMPL
   pip install -e .

2. Copy and paste reproduce scripts to your source directory.

   cp -r reproduce /project/home/your/source/directory

3. Run & modify the scripts

   cd /project/home/your/source/directory
   python simpl_meta_train.py  --help

Downloading SPiRL Pre-trained Model

Currently, we don't support SPiRL skill extraction in this repository. Here are pre-trained SPiRL for each domain :

• Maze
• Kitchen

Download and feed to SiMPL & SPiRL scripts by --spirl-pretrained-path or -s.