Cerebellar Dynamic State Estimation for a Biomorphic Robot Arm

摘要

The cerebellum has been called the brain's "engine of agility". This paper presents a cerebellum-inspired neural network that performs dynamic state estimation and predictive control. The model combines two types of learning within a radial basis function network. Its performance was demonstrated on a 2-link robot arm built with antagonistic pairs of McKibben air muscles. The arm has a gripper end effector to hold and throw a tennis ball. Trajectory data was collected during multiple throwing trials and used to train the model offline. The data were projected onto 2-dimensional state space maps, from which the network learns to estimate state variables and decision boundaries. It successfully learned to trigger the grip release at the proper state for the ball to hit a target. This algorithm should generalize to benefit a wide variety of biomorphic robots.