Modeling of rider-bicycle interactions with learned dynamics on constrained embedding manifolds

Modeling and control of physical human-machine interactions (pHMI) are challenging due to the high-dimensional movement of human body. In this paper, we present a hybrid statistical/physical dynamic model scheme to capture the pHMI through a rider-bicycle interaction example. We use the Gaussian process dynamical model (GPDM) to capture the high-dimensional human movement into a low-dimensional latent space. We extend the GPDM by incorporating additional physical control inputs into the model. The GPDM control inputs are coupled with the physical dynamic model from the bicycle systems such as crank angle etc. The proposed statistical/physical dynamic model is further enhanced by constrained manifold learning algorithms so that we can use less training data sets to obtain the more accurate model. We illustrate the modeling scheme through a lower-limb pedaling example in human bicycling experiments.