Research On The Feeding Tasks-oriented Learning Method Of Service Robot Manipulation Skills

Learn from Demonstration is a research hotspot in the field of robot manipulation skill learning in recent years.Through learning the effective information of the original skill demonstration,the robot gains the ability to reproduce the manipulation skill in the original or new scene.This paper focuses on the skill scene containing actual manipulation objects detected by RGB-D vision sensor,and the robot assisted feeding is taken as the research goal.The robot's manipulation skill learning method is systematically studied from four aspects of skill demonstration,skill segmentation,primitives' representation and skill reproduction.The experimental results of the method proposed in this paper were also presented in a real feeding scene.On the basis of summarizing the research status of service robot manipulation methods and robot skill learning,this paper focuses on the following three aspects:(1)A feeding task-oriented service robot operation experimental platform was constructed based on Kinova manipulator.Under the experiment platform,food detection was completed and established an eye in hand visual system,at the same time,the coordinate system of robot,camera,food and experimenter were transformed into robot base coordinate system,which provides the platform basis for developing the demonstration of robot assisted feeding manipulation skills and data acquisition.(2)Realizing the segmentation of robot assisted feeding manipulation skills based on proposed Bayesian non-parametric model BP-HSMM.The number of features activated in the Beta process is used as the number of hidden states of HSMM for model training,which overcomes the problem that the number of hidden states was artificially specified in the training of general HSMM model,then,the distribution of hidden states in the model is deduced in detail by using BP-HSMM Viterbi algorithm.Finally,the segmentation effect of the model proposed is verified by comparison on the relevant feeding skill demonstration data,which provides an important data source for the subsequent characterization of manipulation primitives and the reproduction of manipulation skills.(3)The robot assisted feeding manipulation skills are reproduced and generalized based on DMP and Behavior Tree.The Dynamic Movement Primitive algorithm adopted in this paper tracks the trajectory when the manipulation primitive changes the starting point,destination and obstacles while retaining the movement trend of the original manipulation primitive.The Behavior Tree framework has clear logic and is easy to be extended.By using the control nodes and action nodes in the framework,a concrete tree is built to realize the reproduction and generalization of robot manipulation skills.In the feeding scene of changing the location of food,the height of the mouth and avoiding obstacles,it is proved that the control method of DMP combined with Behavior Tree is feasible in the field of reproducing service robot manipulation skill.According to the above theoretical and experimental research,this paper designed an experimental system which is based on Kinova-Jaco2 service manipulator and aimed at completing assisted feeding.Correspondingly,this thesis explored specific processes such as skill segmentation,primitives' representation and skill reproduction as well.The validity of the proposed theory and method in the field of service robot manipulation skill learning is verified in two experimental scenarios of feeding bread and drinking water.