A Systematic Review of Deep Reinforcement Learning for Legged Robot Locomotion

Legged robot locomotion remains a critical challenge in robotics, demanding control strategies that are not only dynamically stable and robust but also capable of adapting to complex and changing environments. deep reinforcement learning (DRL) has recently emerged as a powerful approach to automatically generate motion control policies by learning from interactions with simulated or real environments. This study provides a systematic overview of DRL applications in legged robot control, emphasizing experimental platforms, measurement techniques, and benchmarking practices. Following PRISMA guidelines, 27 peer-reviewed studies published between 2018 and 2025 were analyzed, covering model-free, model-based, hierarchical, and hybrid DRL frameworks. Our findings reveal that reward shaping, policy representation, and training stability significantly influence control performance, while domain randomization and dynamic adaptation methods are essential for bridging the simulation-to-real-world gap. In addition, this review highlights instrumentation approaches for evaluating algorithm effectiveness, offering insights into sample efficiency, energy management, and safe deployment. The results aim to guide the development of reproducible and experimentally validated DRL-based control systems for legged robots.