Biped Robot Data-driven Gait Trajectory Genesis for Traipse Ground Conditions

This paper presents a data-driven gait model for continuous parameterization of joint kinematics which yields the genesis of biped robot trajectory. This work employed data-driven approaches such as Deep Neural Network (DNN) and Long Short Term Memory (LSTM) for parameterization using the human locomotion data-set which consists of 10-able subjects walking data on varying inclines and speeds. It allows a smooth and non-switching prediction surface which provides the reference gait trajectory. Additionally, to constrain the model from following the high variance points from the mean trajectory, a loss function that incorporates the standard error of the inter-subject mean is also proposed. Performance evaluation shows that the LSTM performs far better than the DNN in terms of mean and max error for both trained and untrained data-set. Finally, the impact of varying speeds with an incline on the predicted kinematic trajectory for both models is also presented.