Safe reinforcement learning and its applications in robotics:A survey
ZHANG Chang-xin
ZHANG Xing-long
XU Xin
LU Yang
Abstract：Reinforcement learning is a kind of machine learning method that realizes sequential optimization decisions by interacting with the environment.It has been applied in games,recommendation systems and natural language process-ing.However,it is still a challenge to ensure the safety of reinforcement learning algorithms when applied to robotics in the real world.In recent years,the safe reinforcement learning methods for robotics systems have become a hot research direction,gaining extensive attention in robotics and reinforcement learning communities.This paper surveys important achievements and development tendency of safe reinforcement learning based on the existing work and focuses on their applicability in robotics.This paper first introduces the general problem description of safe reinforcement learning.Then we focus on the latest significant progress in this field from the perspective of method and performance,including con-straint policy optimization,control barrier function,safety filter and adversarial training methods,and their applications in autonomous driving vehicles,unmanned aerial vehicles and other robotic systems.Finally,the future research direction of this field is prospected and discussed.
Keywords：roboticssafe reinforcement learningconstrained Markov decision processrobustness
Publication Date：2023-12-28
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：14( 2090-2103 )