Machine Learning for Bipedal Walking

Bipedal robotic locomotion presents a significant challenge to the controls designer. The equations of motion governing these systems are generally hybrid or switched due to intermittent ground contact and consist of numerous coupled non-linear differential equations even in the simplest case. These attributes make traditional control techniques difficult to apply. In this paper, an alternative controller for a 5-link planar biped robot is created through a combination of feedforward neural networks, genetic algorithms and traditional PD control. The neural network uses certain state variables as input and generates a desired target joint state based on the current state in a manner qualitatively similar to HZD. A PD controller than attempts to force the robot into this configuration. In this way the neural network specifies a time invariant trajectory as a function of some combination of state variables. A modified genetic algorithm is used to evolve successful neural controllers for the system.