Abstract:

The need of robots in industries or as the service robots grows significantly in recent decades. Recently, many researches are devoted in the human-robot interactions. For examples, the collaborative working, the service robots for touring or attentive home care, or the exoskeletons for augmenting human powers or assisting the patient for normal functionalities are of interests. In the human-robot interaction, the modeling and the control of the dynamics are essential. With better modeling of the dynamical system, the robot can sensor more and provide accurate and precise responses, whereas the control scheme ensures the safety robustly and let the robot react in a human-friendly way. For better development and the control of the robot in the next generation, this thesis is devoted to the robot dynamics learning and the control of the human-robot interaction with applications in the three topics: the learning of the robot dynamics with the structured kernels, the virtual impedance control for safe human-robot interaction, and the Bayesian exoskeleton.

To model the robot dynamics automatically and accurately, the structured kernel concerns the system identification with the machine learning techniques. It is well-known that the proper model of the robot can boost the performance of the control. The tradition parametric models based on the analytic formulation are often too complex for general systems, while the other methods such as the autoregressive-moving-average (ARMA) model or the general machine learning models introduce either bias or variance in learning. The interest here is how to design a general system identification scheme that enjoys the benefits of both. The proposed kernels are designed so that not only the structures of the analytic model is implicitly modeled but the system can be identified as the blackbox as in the machine learning approaches. In short, the proposed method is a system identification framework that learns the system automatically without any derivation of the system dynamics and yet converges to analytic model pointwisely as the number observations goes to infinity. For the control of the human-robot interaction, the virtual impedance control and the Bayesian exoskeleton system are proposed. For the robots with individual functionalities, the virtual impedance control is designed for the robust, smooth, and consistent collision avoidance that the robot can avoid all the possible collisions robustly while trying to accomplish the original task. Therefore, the robots can response the human nearby safely and compliantly. As for the robots on the human body, the Bayesian exoskeleton system assists human operators with the robust hybrid control. In the Bayesian exoskeleton system, the Bayesian estimator inferences the human intention adaptively, and the inner assistive torque control can ensure the robustness of the system by considering the ability of the operator. Therefore, the resultant exoskeleton system can ensure both the safety and the effective assistance.