Abstract: With the widespread adoption of articulated manipulators in industrial manufacturing, maintenance, and repair operations, motion planning has garnered significant attention as an essential functionality. Compared to classical motion planning techniques, such as artificial potential fields, sampling-based methods, and heuristic algorithms, which often suffer from high computational demands and low convergence rates, reinforcement learning (RL) offers compelling advantages. Its inherent efficiency, generalizability, and capability in handling complex problems have positioned RL as a prominent research focus within the motion planning domain. While numerous RL-based motion planning algorithms have emerged, a systematic classification of existing contributions and a precise evaluation of optimization strategies for planning metrics remain crucial. Firstly, we introduce the review methodology and search strategy and methods for relevant technologies. Subsequently, we comprehensively analyze how RL and fusion algorithms address key performance metrics, including high-dimensional space computation, safety-performance trade-offs, and sim-to-real transfer, across both structured and unstructured environments. Finally, we discuss the latest advancements and future research directions that will specifically address emerging needs in unstructured environments.