An integrated physical-learning model of physical human-robot interactions with application to pose estimation in bikebot riding

Abstract

Modeling physical human-robot interactions (pHRI) is important in studying human sensorimotor skills and designing human assistive and rehabilitation systems. One of the main challenges for modeling pHRI is the high dimensionality and complexity of human motion and its interactions with robots and the environment. We present an integrated physical-learning pHRI modeling framework with applications to the bikebot riding example. The modeling framework contains an integrated machine-learning-based model for high-dimensional limb motion with a physical-principle-based dynamic model for the human trunk and an interacted bicycle-like robot (bikebot). A new axial linear embedding algorithm is used to obtain the low-dimensional latent dynamics for highly redundant human limb movement. The integrated physical-learning model is then used to estimate human motion through an extended Kalman filter design without using any sensors attached to the limb segments. Extensive bikebot riding experiments are conducted to validate and demonstrate the integrated pHRI model. Comparison results with other machine-learning-based models are also presented to demonstrate the superior performance of the proposed modeling framework for bikebot riding.