Traditional humanoid robots using rotary joints face issues such as high motor force demands during sustained standing, resulting in overheating during prolonged operation. To address these challenges, we propose a humanoid robot equipped with planetary roller screw actuators at the knee joints. A detailed kinematic model is developed to accurately relate linear actuator displacement to knee joint rotation, providing the foundation for effective control. Building on this model, we train a reinforcement learning policy using the Proximal Policy Optimization (PPO) algorithm within the NVIDIA Isaac Gym environment. The control architecture adopts an asymmetric actor-critic design, leveraging privileged information during training to improve learning efficiency. The trained policy is first validated in simulation using MuJoCo and then deployed on real hardware. To support real-world operation, a hybrid control framework is implemented, combining torque-based control for rotary joints with position control for the roller screw-driven knee joints. Experimental results confirm the robot’s ability to maintain stable, accurate, and consistent gait performance over extended periods.