Reinforcement Learning for Quadrupedal Robot Control: Taxonomy, Sim-to-Real, Robustness, and Emerging Trends

Quadrupedal robots exhibit strong mobility in complex environments where wheeled platforms often perform poorly, but their control remains difficult. In recent years, reinforcement learning (RL) has received growing attention in quadrupedal locomotion, as it supports direct policy optimization without relying entirely on hand-crafted control rules. This paper presents a control-oriented review of RL-based quadrupedal robot control. It first introduces the main physical and algorithmic foundations. Then it examines recent research progress from three complementary aspects: control architecture, observation and information design, and training task formulation. In addition, it analyses several persistent obstacles to practical deployment, namely sim-to-real transfer, reward and objective design, real-time robustness with online adaptation, and perception–actuation integration. The results suggest that future progress will depend not only on stronger learning algorithms, but also on more transparent evaluation, more hardware-aware system design, and more reliable integration of perception, dynamics, and control.