Training and Transfer of a PID Balance Controller for Quadruped Robots Using Artificial Neuronal Networks

Maintaining balance in quadruped robots requires precise coordination of joint movements, which varies depending on the unique physical characteristics of each robot, such as dimensions, mass distribution and centre of gravity. There are several control methods for quadruped robots balance, a commonly used is the Proportional Integral Derivative (PID) controllers, provide robust stability, but typically require individualised tuning for each robot due to these varying physical parameters. To address this limitation, this paper explores an intelligent control strategy that leverages neural networks to generalise balance control across different quadruped platforms. Initially, a PID controller was implemented to create a large dataset by controlling the equilibrium of a commercial 12 joints quadruped robot. This data is used to train a several perceptron neural networks to learn the complex mapping of body orientation to joint movements. Through a parameter search, it was determined that a simple single-layer neural network with 18 neurons effectively mimicking the behaviour of the PID controller. This neural network is then applied to a secondary quadruped robot with different dimensions and mass, demonstrating that single-layer networks, despite their simplicity, can effectively capture essential control dynamics, reducing model complexity and enabling rapid deployment on different quadruped robots. Furthermore, this work opens the way to scalable and adaptive control methods in robotic systems where neural networks trained on one platform can be effectively transferred to others with minimal modifications.