Unsupervised learning of quadruped robot balance

This paper reports a research activity grounded on year-long experience in unsupervised neural networks used in a lot of different scenarios and here applied to control balance in a quadrupedal robot structure. Simulation results are reported on balance control of a mini cheetah quadruped robot undergoing unexpected weight disturbances coming from additional loads applied on the robot body. The unsupervised learning structure, belonging to the family of Motor Maps controllers, is shown to rapidly learn the additional torques to be applied to the robot legs in order to compensate for the disturbance and so contributing to improve the underlying Whole Body Impulse Controller in front of the robot model uncertainties.