A Method Of Robot Imitation Learning Based On Structural Grammar

In the process of human growth,we can acquire operational skills through perception,imitation and learning.With the development of artificial intelligence and computer vision technology,people also begin to consider whether robots can perceive,understand and generalize tasks from the environment just like humans to deal with complex and diverse operation scenarios.According to the above content,this paper carries out the research of robot imitation learning based on structural grammar.It mainly includes the following aspects:Firstly,a calibrated Kinect V2 camera was used to capture and align the color and depth maps of the demonstration activity.The Open Pose algorithm was used to identify the two-dimensional bone joint coordinates in the color image,and the 3D body posture information sequence was obtained by combining with the depth map.A feature description method based on central node was selected to extract the pose features of human upper limbs.Form the human body posture characteristic sequence of demonstration activities.Secondly,to solve the problem of behavior segmentation,a behavior segmentation method based on Cam Shift algorithm is proposed.Taking advantage of the characteristic that the demonstration action and the object involved always appear at the same time,the area near the hand joint is extracted according to the two-dimensional joint coordinates obtained from Open Pose,and the object near the hand is tracked by Cam Shift algorithm to achieve the purpose of behavior segmentation.At the same time,the object information involved in each action after segmentation is obtained.In the aspect of action recognition,a hybrid training model based on AIA and BW was proposed based on the characteristics of Artificial Immune Algorithm(AIA)and the problems existing in the traditional HMM training method Baum-Welch(BW).The proposed model was used to extract the action information of the demonstration activity.Combine the above two methods to transform the Hanoi tower demonstration activity into {Action-Object} sequence.Thirdly,the proposed method focuses on the learning problem of demonstration strategy and proposes a learning method of demonstration strategy based on structural grammar.The{action-object} sequence was abstracted into action primitive sequence,and the Probabilistic Context Free Grammar(PCFG)method in structural Grammar was used to characterize the action primitives sequence,construct the initial Grammar,and calculate the rule probability.At the same time,the subset probability set of the sequence of action primitives is obtained.The Chunk operation and Merge operation are used to transform the initial grammar to form the grammar space.The Minimum description length(MDL)criterion was used to evaluate the grammars,and the Beam Search method was improved to find the optimal grammars in the grammars space,that is,the strategy of the demonstration activity,which was used to guide the robot to perform more complex higher-order tasks.Finally,the data synthesis experiment and the Hanoi tower experiment were carried out.In the data synthesis experiment,compared with MDL data of other methods,it is verified that this method has good data compression performance and strong anti-interference ability.In the experiment of Hanoi tower,several groups of third-order Hanoi tower demonstration activities is obtained.The strategy is used to observe the fourth-order Hanoi tower demonstration activities,and then guide the robot to perform the fourth-order Hanoi tower task.Experimental results show that this method has strong generalization ability and good anti-interference performance.