This repository contains Rainbow DDPG algorithm from paper Sim-to-Real Reinforcement Learning for Deformable Object Manipulation along with a toy pushing task to demonstrate how to use the code.

The code was tested on Mac OS with Python3.6. Use of virtualenvs is recommended. To run:

pip install -r requirements.txt
python main.py

Runnign a full training may take more than 24 hours on a machine with Nvidia Titan GPU and use a considerable amount of memory.

To run a demonstration of the toy task:

pip install -r requirements.txt
python run_demo.py

Please note that the hyper parameters are not necessarily optimised for the task.

For a complete list of references, please see the accompanying paper.

The learning algorithm is based on OpenAI baselines (https://github.com/openai/baselines), the perlin noise file is heavily based on https://github.com/nikagra/python-noise/blob/master/noise.py and robot meshes are generated from https://github.com/Kinovarobotics/kinova-ros.