Human-Robot Interaction Based On Imitation Learning And Hidden Markov Model

Recently, as the development of the robot industry, more and more robots will walking into our life from the lab and factory. Frequently interaction bring new requirement to the robot. Robot should have adequate adaptation ability to adapt a changing environment. Besides, robots’ motion should imitate human motion in a way that human feels nature and comfort. In the past, and still to a large extent now, robots had to be tediously hand programmed for every task they performed. But when robots are used in our daily life, they need mostly conducting interactions with non-specialist humans. It’s hard to imagine how to anticipate all th e missions robot would face in daily life, and writing programs for them. This require the robot to have the ability that learn a new movement in a simple way, and the learned movement should be robust and adaptability. According to the above questions existed in human robot interface, in this dissertation, I conduct research based on imitation learning and Hidden Markov Model.In this dissertation we used the Hidden Markov Model to model human gesture sequence and identify human intention through human gesture. An efficient while simple quantize algorithm was proposed to convert human gesture into observations. Then Baum-Welch algorithm was implemented to learn model parameter. Furthermore this dissertation introduced a threshold model to distinguish pre-defined gestures from non-defined gestures. When user conduct movements, in this dissertation we used forward-backward algorithm to calculate the probability of current observation under every learned models.Traditional research on imitation learning based human robot interaction mostly used the DMPs(Dynamic Motor Primitives) method. DMPs method could maintain high efficiency for robot with multi-DOF. But DMPs method could not handle the temporal perturbations. Moreover, DMPs could not learning a generaliz ed dynamics from multiple demonstrations. In this dissertation we assume that human would expect different kinds of interaction movement from different interaction space. Based on this assumption, in this dissertation we investigate daily life human-robot interaction problem. Focuses the problem in the DMPs method, we implemented SEDS(Stable Nonlinear Dynamical Systems) method to learn movement from human demonstrations. SEDS method build a nonlinear dynamic system based on Gaussian Mixture Models. When certain constrains are satisfied the dynamic system is asymptotically global stable. The parameters of the Gaussian Mixture Models can be estimated using nonlinear optimization method combined with an EM algorithm.To test the feasibility of our method, in this dissertation we design an experiment to simulate everyday contact between human and robot, ask the robot to conduct high-five and hand shake action with human. Based on the high-five and hand shake demonstration trajectories, in this dissertation we us e MATLAB simulation to choose appropriate K and objective function, and use theory to explain why we choose such objective function. The effectiveness of our method are tested by the experiment result. The result of this dissertation could be implemented o n the interaction between service robots and human, furthermore, the result can be used in the factory for passing tools.