The sensor data measured by a quadruped robot plays a crucial role in recognizing the robot’s physical environment and controlling its posture. High accuracy in sensor data is essential for precise robot control. However, most high-performance sensors are expensive and often subjected to excessive loads during field operations.

This project explores virtual sensor technology as a viable solution. The goal is to reduce dependency on high-cost, high-maintenance sensors by using a virtual sensor system trained with actual sensor data combined with hardware information, enabling the substitution of physical sensors with virtual ones.

The approach employs neural network theory with a focus on back propagation.