Research On Object Grasping And Active Security Technology Of Collaborative Robots For Indoor Service Scenarios

Collaborative robots for indoor service scenarios usually need to complete the task of picking and delivering objects.However,the unstructured,complex and variable characteristics of indoor scenes pose great challenges for the collaborative robots to accurately grasp objects.In addition,for some human-robot collaboration tasks,human-robot security is paramount.At present,the research on object grasping and active security of cooperative robots still uses the processing mode of structured scene of industrial robots.This model does not take into account the effects of complex background,object occlusion,illumination change,etc.,making it difficult to meet the needs of indoor service scenarios.Therefore,this thesis studies and proposes a set of solutions for object grabbing and active security of collaborative robots facing indoor service scenarios.Aiming at the problem of object recognition,this thesis proposes an improved Point Net network for object cluster recognition,which uses a symmetric function to overcome the unstructured and disordered characteristics of point cloud.The improved Point Net simplifies the structure of Point Net and speeds up the development of the model.Compared with the template matching method using a feature descriptor,it achieves higher recognition accuracy.Aiming at the problem of object grasp detection,this thesis studies the local sampling method based on point cloud to generate candidate grasps,develops a grasp evaluator to screen out the best candidate grasp,and finally converts it to the robot coordinate system by hand-eye matrix.This grasp detection method can generate available grasp configurations by sampling less candidate grasps,and can be used to grasp new objects.Aiming at the problem of robot active security,this thesis proposes a human detection algorithm combined with the extraction of human point cloud clusters,which uses 3D information of point cloud to avoid the trouble of human body positioning.The algorithm is robust to the situation of human occlusion and multiple people approaching each other in indoor scenes.Based on the distance test between the robot and human bounding box,the robot achieves active collision avoidance in complex indoor scenes.In order to verify the feasibility of the improved Point Net network,the local grasp detection method based on point cloud,and the human detection algorithm combined with point cloud,this thesis designs three experiments based on Baxter collaborative robot: designated object grasping,multi-object cleaning,and human-robot collision avoidance.The experimental results show that the proposed method or algorithm can enable the robot to accurately recognize the target object,find the appropriate grasping point and actively avoid collision when the person approaches it.