Deep Reinforcement Learning for Dynamic Obstacle Avoidance of Mobile Robots in Indoor Environments: A Review

Abstract

Highlights What are the main findings? Systematically classifies DRL algorithms for indoor mobile robot dynamic obstacle avoidance. Analyzes core challenges and improved strategies across five key technical dimensions. What are the implications of the main findings? Provides a unified reference framework for robot navigation and DRL algorithm selection. Clarifies promising future directions for indoor dynamic obstacle avoidance research. Abstract The ability of mobile robots to avoid obstacles dynamically in indoor environments is a necessary condition for achieving autonomous planning and navigation. When dealing with unstructured and randomly dynamic indoor scenes, traditional obstacle avoidance algorithms have poor adaptability and low flexibility, making it difficult to handle environmental uncertainties. Deep Reinforcement Learning (DRL), with its efficient end-to-end decision-making, autonomous interactive learning capabilities, and proficiency in modeling complex dynamic systems, has emerged as a focal point of research in dynamic obstacle avoidance. This paper first presents the theoretical foundation of DRL, then categorizes the fundamental DRL algorithms for indoor dynamic obstacle avoidance into three main types: Value function-based, Policy-based, and Actor-Critic-based algorithms, while also introducing relevant algorithms. Furthermore, it addresses the core challenges encountered in indoor dynamic obstacle avoidance and summarizes various improvement strategies for the different basic algorithms, detailing their starting points and performance impacts. Finally, the paper outlines the development trends and future research directions in this domain. This review aims to serve as a systematic reference for the design and engineering application of DRL algorithms in dynamic obstacle avoidance for indoor mobile robots.