Reinforcement learning, an influential technique for training intelligent systems, has gained prominence in academia and industry due to its ability to solve complex problems without pre-existing models. However, its application to real-world systems is challenging due to its complexity caused by high non-linearities, limited samples, and constraints. Consequently, research in this area has mainly focused on simulation, where models can be easily tested and refined. In this work, a reinforcement learning strategy towards a real-world bipedal robot to learn walking behavior from scratch was proposed. We present a robot system design focused on stress reduction and simplicity, ensuring robust performance, including a custom printed circuit board (PCB) for efficient handling of electrical components. The system s software is divided into the master system board and the ROS system, allowing communication between components and addressing data loss and communication delay issues. The robot model simulation is developed on the Mujoco platform, incorporating physical properties and environmental parameters. We utilize Deep Deterministic Policy Gradient (DDPG), Twin-Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC) algorithms for learning and evaluating the policy distillation technique for transferring knowledge to a more efficient network. Finally, we evaluate the transfer of learning to the real world and present a preliminary experiment of learning from scratch in the real world. Our results demonstrate the effectiveness of the robotic system design and the learning algorithms, achieving stable walking in simulation and a maximum of fourteen steps in real life with the policy distillation of the SAC algorithm.