Human-Robot Collaboration Based on Gaussian-Mixture Model

A method based on Gaussian Mixture Model(GMM) is proposed for learning the relationship between image observation and robot joint angles. When robot and human are working together, observation variables are acquired through camera image and GMM is the used to compute the generative probability of that observation. The generative probabilities of different types are unified and the type which has biggest posterior probability is selected as the identification result. And the Gaussian Process of that type is used to generate a group of joint angles to control the movement of robot, which can adapt to the change of human movement. The whole process is proceeded in a Bayesian mechanism, which can reduce the need for training sample and improve system performance. And the experiments show that GMM can approximate the non-linear relationships, which makes robot learn directly from demonstrations, and thus eliminates the need for camera calibration as well as inverse kinematics of robot. The proposed method can build an learning model in limited demonstration data set and is suitable for modeling multiple collaborative patterns in different space zones.