Research On Demonstration Learning Method Of Compliant Assembly Skills Based On 3D Vision

Assembly is an important part of production and manufacturing.Due to the collision contact and pose uncertainty in the assembly process,once the movement is not accurate,it will lead to the consequences of light failure and heavy damage.Therefore,robot assembly operation relies heavily on experts’ manual programming,which is time-consuming and laborconsuming.Inspired from the learning process of teaching by words and deeds between people,it is of great significance to study how the robot can quickly and accurately carry out compliant assembly like human beings by referring to the characteristics that human beings can quickly master an assembly skill and flexibly face the unpredictable situation in the assembly process.Learning from demonstration(LFD)is a new way of knowledge transfer,which can make the robot system have the ability to extract and understand the effective information from the operator’s demonstration,and then transform the information into the robot control program and related parameters.Compared with the simple pre programming method,learning from demonstration provides a very flexible means of robot programming.Demonstration learning is an effective way to learn flexible assembly skills.However,the existing demonstration learning methods usually focus on robots and people,which will bring some problems,such as lack of key information,limited movement,physical contact risk,serious ontology coupling and so on.In addition,demonstration learning relies too much on professional equipment,regards flexible assembly skills as a black box problem,lacks data analysis between each other,relies on multi-modal information,so it is difficult to demonstrate,and the learning efficiency is low,and it is difficult to effectively express complex flexible assembly skills by expressing robot motion with traj ectory.Therefore,in this paper,3D vision is used to obtain the pose data of the workpiece directly for demonstration learning.The demonstration process is separated from the robot system,and the assembly movement is separated from human and robot,and the core assembly skills are directly learned by object-oriented.For the learning of compliant assembly skills,this paper analyzes the causes of compliant assembly skills,and uses single-mode data to restore multimode information,combined with contact features,constraint features and motion features to solve the problem of expression and learning of compliant assembly skills.The main research work of this paper is as follows:First of all,aiming at the demonstration data acquisition,this paper proposes an object pose recognition algorithm,and uses QT to develop a real-time pose recognition software;Aiming at the assembly error caused by the uncertainty of demonstration data,this paper proposes a particle based attitude optimization algorithm for the spatiotemporal consistency of contact state.The main process is as follows:developing virtual contact simulator based on open cascade to identify the contact state;Aiming at the shortcomings of large detection error and low measurement accuracy of single sensor,a pose optimization algorithm based on Kalman filter is proposed;Based on the idea of particle,the pose of the model is evaluated,and the probability model of the contact state under the virtual contact simulator is established,so as to correct the attitude uncertainty of the assembly;According to the principle of time-space consistency of contact state,the segmentation point is confirmed by sliding window and hidden Markov model,and the final pose optimization is carried out based on the velocity.Secondly,for the learning stage,the parts are often accompanied by physical contact in the assembly process,and the contact produces constraints,which determine the external movement of the object.The contact constraint movement interaction is causal,and the assembly skills are coded in the constantly changing contact point set,which constitute a macro description of continuous movement.Therefore,starting from the set of contact points,this paper proposes a method of generating compliance control strategy based on the set of contact points to realize the micro analysis and decoding of macro demonstration motion.For assembly,it often involves multiple subtasks,and each subtask is regarded as a skill.Therefore,this paper proposes to segment the assembly task based on the set of contact points.The algorithm obtains the set of contact points by establishing the boundary model of contact,and then models the set of contact points to realize the identification of static contact conditions.The subtask sequence is obtained by analyzing the dynamic set of contact points to realize the subtask segmentation.In order to make the robot have the ability of compliant assembly,the flexible control strategy of the divided sub task is learned.In this paper,the local feature analysis and virtual work theorem are used to obtain the contact features,and the switching conditions between different strategies are learned by combining the position information.For constrained motion,the velocity constraint equation is established and the corresponding selection matrix is obtained to realize force position decoupling.For unconstrained motion,the DMP is used to learn,and the compliance control strategy is generated by combining the learning results.Finally,in order to verify the correctness and effectiveness of the compliant assembly skill control strategy proposed in this paper,a virtual simulation environment is constructed based on V-REP,and preliminary verification is carried out for the shaft hole assembly,and physical experiments are carried out based on the ABB IRB1200 robot.The experimental results all prove the accuracy and effectiveness of the compliant assembly strategy based on demonstration learning proposed in this paper.