ADVANCES IN DEEP REINFORCEMENT LEARNING FOR AUTONOMOUS MOBILE ROBOT NAVIGATION: A SYSTEMATIC REVIEW OF ALGORITHMS, SIM2REAL, AND SAFETY

Abstract

Despite decades of progress, navigating autonomous mobile robots (AMRs) through stochastic and non-deterministic environments remains a formidable challenge. While classical geometric-based planners provide rigorous theoretical guarantees, their performance often degrades under high-dimensional uncertainty. This review dissects the paradigm shift toward Deep Reinforcement Learning (DRL), emphasizing its capacity for end-to-end perception-to-action mapping. We categorize contemporary DRL architectures into value-based, policy-gradient, and actor-critic lineages, evaluating their efficiency in complex workspaces (Haarnoja et al., 2018; Hasselt et al., 2015; Schulman et al., 2017). Crucially, we scrutinize the field's persistent "pain points": training sample inefficiency, the notorious Sim2Real gap (Da et al., 2025; Loquercio, 2023), and the exigency for verifiable safety constraints (Gu et al., 2024). By synthesizing landmark studies from 2015 to 2025, this paper contributes a novel taxonomy and explores the emerging trajectory of Foundation Model-driven navigation, providing a roadmap for the next generation of socially-aware and resilient AMR systems.