Resumen

This thesis presents a vision-based framework for real time human motion imitation on a humanoid robot, validated on the Unitree G1 platform. Using a single RGB-D camera and a BlazePose based pipeline, three dimensional human arm motion is captured and mapped to the robot without wearable sensors or predefined motion templates. Shoulder and elbow joint targets are computed through geometric, vector based kinematic relationships and transferred to the humanoid as angles using a task space inverse kinematics formulation that minimizes elbow and wrist position errors. The system is first validated in a MuJoCo simulation environment with a detailed model of the Unitree G1, enabling safe tuning before real robot deployment. To ensure smooth and stable motion, a first order low pass filter is applied to the commanded joint trajectories. Experimental results in both simulation and hardware demonstrate accurate and stable replication of human arm movements. The proposed framework shows that vision-based imitation is a feasible and flexible approach for humanoid control, and serves as a foundation for future extensions such as full body imitation, learning based methods, and interactive human robot applications.