Research On Robot Grasping Trajectory Planning Based On Vision And Dynamic Movement Primitives

With the rapid development of automation technology and artificial intelligence,robot technology has been widely used in various industries such as industrial production and service fields.The flexibility and complexity of robots are also increasing.For robots engaged in various tasks,the operating environment and task goals are usually complex and changeable.Simple and repeated pre-programmed instructions can no longer meet the needs of actual work tasks.Having certain self-learning ability has become a necessity for robot skills.Learning from demonstration is a way for robots to learn motor skills through human demonstrations.It eliminates repetitive programming processes and enables robots to quickly reproduce teaching actions in new environments.However,the existing teaching and learning methods have certain limitations.The robot has a single way to obtain the position of the target point,and the trajectory planning method lacks autonomy.As a result,the robot cannot quickly and flexibly adapt to the new environment.To solve these problems,this paper proposes a trajectory planning method based on vision and dynamic movement primitives for the grasping task.The method is divided into three stages,which are optimal demo trajectory extraction,trajectory learning and object recognition grasping.In the first stage,due to the high quality requirements of the training model for teaching learning,the most perfect demonstration trajectory samples cannot be given at one time during manual teaching,and there may be multiple jitters during the dragging of the robotic arm.Therefore,this thesis proposes a new method to obtain the optimal teaching trajectory.Firstly,manually collect multiple teaching trajectory samples,and smooth the Gaussian filtering process for each sample.Furthermore,use the Gaussian mixture model to multiple trajectory samples perform clustering representation to extract the features of multiple teaching trajectories.Finally,generalize the teaching trajectories by Gaussian mixture regression to generate the optimal teaching trajectories for specific operation tasks.In the second stage,a dynamic motion primitive algorithm is used to learn the optimal teaching trajectory obtained in the previous step,extract the characteristics of the teaching trajectory,and use local weighted regression to obtain the weighting parameters of the teaching trajectory.In the third stage,in order to enhance the autonomy and flexibility of trajectory planning,this thesis integrates the target detection algorithm based on deep learning into the trajectory planning method.Firstly,the visual perception of the environment is captured through the Kinect depth camera,and then using the Mask R-CNN algorithm to recognize and locate the object to be grasped.Furthermore,input the three-dimensional space position of the grasped object into the second stage dynamic motion primitive algorithm after coordinating transformation.Finally,calculate the forcing function of the learning trajectory based on the weight parameters obtained in the previous stage According to the new target position and the forcing function,a new grasping trajectory is generalized and transmit it to the robot for execution.This thesis builds an experimental platform based on a robot operating system,establishes a Kinect vision system and performs parameter calibration,object detection experiments and trajectory learning experiments with different degrees of freedom,and verifies the trajectory planning method proposed in this paper based on the existing Baxter robot in the laboratory Feasibility.The experimental results show that the robot can learn the movement trend of the teaching trajectory,and can automatically identify the grab target and complete the grab task,which proves the feasibility and effectiveness of the method in this thesis.