This paper presents the design and development of a quadruped robot engineered to learn autonomous locomotion through reinforcement learning (RL) techniques. A core innovation involves the creation of a physical robot prototype (called “Xiinbal”) along with its high-fidelity digital twin within the MuJoCo simulator, enabling precise spatial training through virtual environments that accurately reflect the robot’s real-world dimensions and physical properties. The RL algorithms, specifically Proximal Policy Optimization (PPO), are trained to facilitate stable and agile forward locomotion by optimizing joint movement sequences, prioritizing displacement, and maintaining dynamic stability based on linear and angular velocities and orientation (yaw, pitch, roll). This inherent spatial control is further augmented by successful 3D environment reconstruction capabilities, utilizing both 2D camera depth estimation and an Intel RealSense L515 camera for Simultaneous Localization and Mapping (SLAM), it is crucial for advanced spatial understanding. This accessible and modifiable platform holds substantial potential for innovative applications in diverse urban spatial environments, including autonomous inspection and mapping of challenging infrastructures, delivery in complex city terrains, environmental monitoring applications, such as air quality assessment with integrated PM10 and PM 2.5 sensors and thermal inspection, urban structures using exploration where the robot’s enhanced stability and mobility offer considerable advantages.