Modeling Gait Using CPG (Central Pattern Generator) and Neural Network

Abstract

In this study, we utilize CPG (Central Pattern Generator) concept in modeling a bipedal gait. For simplicity, only lower extremity body of a biped is considered and modeled. Actually, gait is a result of a locomotor which is inherent in our bodies. In other words, the locomotor applies appropriate torques to joints to move our bodies and generate gait cycles. Consequently, to overcome the gait modeling problem, we should know structure of locomotor and how it works. Actually, each locomotor mainly consists of two parts: path planning and controlling parts. Task of path planning part is to generate appropriate trajectories of joint angles in order to walk properly. We use CPG to generate these proper trajectories. Our CPG is a combination of several oscillators because of the fact that gait is a periodic or semi-periodic movement and it can be represented as sinusoidal oscillators using Fourier transform. Second part is to design a controller for tracking above-mentioned trajectories. We utilize Neural Networks (NNs) as controllers which can learn inverse model of the biped. In comparison with traditional PDs, NNs have some benefits such as: nonlinearity and adjusting weights is so much faster, easier and fully automatically. Lastly, to do this, someone is asked to walk on a treadmill. Trajectories are recorded and collected by six cameras and CPG can then be computed by Fourier transform. Next, Neural Networks will be trained in order to use as controllers.