VinceLab401 / SSA-RL

Safe and Sample efficient Reinforcement Learning for Clustered Dynamic Uncertain Environments

Table of Contents

• Introduction
• Install
• Usage
• Acknowledgments

Introduction

We provide code for evaluate the safety and sample-efficiency of our proposed RL framework.

For safety, we use Safe Set Algorithm (SSA).
For efficiency, there are more strategies you can choose:
1, Adapting SSA;
2, Exploration (PSN, RND, None);
3, Learning from SSA;

The video result is shown below, agent is trained to drive to the goal while avoiding dynamic obstacles. The red means SSA is triggered.

Please cite our paper as:

@article{chen2021safe,
  title={Safe and sample-efficient reinforcement learning for clustered dynamic environments},
  author={Chen, Hongyi and Liu, Changliu},
  journal={IEEE Control Systems Letters},
  volume={6},
  pages={1928--1933},
  year={2021},
  publisher={IEEE}
}

Install

conda create -n safe-rl
conda install python=3.7.9
pip install tensorflow==2.2.1
pip install future
pip install keras
pip install matplotlib
pip install gym
pip install cvxopt

Usage

python train.py --display {none, turtle} --explore {none, psn, rnd} --no-qp --no-ssa-buffer
python train.py --display {none, turtle} --explore {none, psn, rnd} --qp --no-ssa-buffer
python train.py --display {none, turtle} --explore {none, psn, rnd} --no-qp --ssa-buffer

• --display can be either none or turtle (visulization).
• --explore specifies the exploration strategy that the robot uses.
• --no-qp means that we use vanilla SSA.
• --qp means that we use adapted SSA.
• --no-ssa-buffer means that we use the default learning.
• --ssa-buffer means that we use the safe learning from SSA demonstrations.

You may also try to test other safe controller (CBF, Shield) by uncommenting line 108-109 and 155-157.

Acknowledgments

Part of the simulation environment code is coming from the course CS 7638: Artificial Intelligence for Robotics in GaTech. We get the permission from the lecturor Jay Summet to use this code for research.