Research On Robot Task Learning Method Based On Visual Demonstration

With the rapid development of robot application field in recent years,the method of automatically mastering new tasks by observing human demonstration has become another research hotspot in robot field.Generally speaking,this field mainly involves two aspects: teaching information acquisition and robot task learning.With the continuous innovation of robot technology,the acquisition method of teaching information has changed from the traditional way based on off-line programming and teaching box to the later dragging teaching mode,and then to the teaching mode based on visual observation at the present stage.In the human demonstration,the recognition and pose estimation of the operated object is the core of visual observation.Different from the general visual perception tasks,the features of the target objects involved in the demonstration are various and the objects are changed frequently.In view of these characteristics,algorithms based on multimodal local feature RGB-D patches have a good application prospect.However,the existing RGB-D patch features are not rotation invariant and sensitive to foreground occlusion and background interference.Because these problems can not be ignored in the demonstration scene,this paper proposes a local feature E-patch with rotation invariance and robustness to environmental interference.E-patch takes the foreground depth edge pixels as the center and samples along the depth gradient direction,so it has good rotation invariance.According to the depth detection results,the foreground occlusion and background interference areas in E-patch are eliminated,so as to improve the robustness to environmental interference.In the framework of Siamese Network,CNN based feature encoder is trained to map E-patch similarity measure to Euclidean distance space of feature vector.Combining the advantages of feature matching and pose voting algorithm,this paper proposes an object detection and pose estimation algorithm based on E-patch feature.In the off-line phase,after reconstructing the mesh model of target object with the Ch Ar Uco calibration board,RGB-D rendering views are obtained from the uniformly distributed sampling perspectives,and the feature codebook is constructed using the E-patch extracted from each rendering view.In the online stage,hypothetical poses are generated by the algorithm framework of feature matching and pose voting.Due to the interference of invisible points in the object model on ICP registration,only visible surfaces are used to verify and refine hypothetical poses.In the aspect of robot task learning,imitation learning methods aim at reproducing human demonstration,but they are usually divorced from the actual situations of robot systems;reinforcement learning methods are oriented to the actual systems,but they face the problem of dimension disaster due to the huge space of action exploration.In this paper,combining the advantages of the two methods,a robot task learning method is proposed,which considers both convenience and reliability.Firstly,the teacher only need to complete a single visual observation oriented demonstration.Then,based on DMPs framework,the robot action exploration space is reduced by imitating and generalizing the teaching information.Finally,DDPG reinforcement learning algorithm is used to optimize the robot action strategy.To improve the success rate of action strategy in training process,the motion controller is designed based on the point attractor system.To verify the effect of this task learning method,an experimental platform for learning pick-place task is built based on UR robot,Kinect depth sensor and host computer.In the experiment,the motion controller of the robot is trained based on the imitation learning and reinforcement learning algorithm,and then the spatial pose of the operated object in the initial scene is estimated based on the visual perception algorithm.Finally,the object pose is substituted into the motion controller and the UR robot is driven to complete the pick-place task.The experimental results show the feasibility and effectiveness of the proposed robot task learning method based on visual demonstration.