Study On The Machine Vision Based Intenngent Grasping For Collaborative Robots

Collaborative robot refers to the robot that can interact with people directly in the working area,which is a promising field of robotics research.Nowadays,the scale of collaborative robot industry in China continues to increase.As an important function for robots,the robotic grasping has been researched for many years and huge progress has been made.But the lack of intelligence in grasping process is still the main problem restricting the further application of collaborative robot in industrial production.The research of collaborative robots' grasping's intelligence can be divided into two sub-question: object recognition,pose estimation and path planning.On the one hand,due to most of the industrial accessories were texture-less,the traditional method cannot be used to object recognition.Meanwhile,scattered stacking,occlusion among each other,as well as the requirement of recognition speed and accuracy,all of which put forward higher requirements for object recognition and pose estimation algorithms.On the other hand,how to plan a feasible path for the robot so that it can complete the movement from the initial state to the goal state,collaborative with people safely and quickly is also a key problem.This thesis is written for collaborative robot's intelligent grasping in the industrial environment,use RGB-D camera as the main sensor,solutions are proposed for training data preparation,semantic segmentation,point cloud registration and path planning algorithms involved in robot grasping,and build the collaborative robot's intelligent grasping hardware and software system.The main contents are as follow:(1)In order to realize object recognition and pose estimation,a convenient automated data acquisition and annotation method was implemented.User only need hand-held RGB-D sensor and move around objects in the scene to record continuous video data in LCM format,and semantic segmentation and pose annotation labels of objects can be generated automatically after simple pose adjustment.This method is used to process,collect and label the scenes with eight objects that included in the Dex-net dataset.(2)Propose a method for object recognition and pose estimation based on the semantic segmentation.First conduct segmentation on color image,then extract object's pointcloud from the depth image.use according to the normal sampling cloud from the nearest point of the corresponding points of iterative algorithm for accurate registration object position.It can effectively estimate the position and pose of texture-less objects with shielding each other in complex environment,providing necessary information for subsequent grasping.(3)The intelligent grasping software and hardware system of the collaborative robots were built,and the above methods were encapsulated into software modules under the modularization thought.And conduct handeye calibration for the whole system.Finally,perform grasping test in the actual environment to verify the effectiveness of the above methods.The dissertation's research on semantic segmentation and pose data labeling,object recognition,pose estimation and path planning could not only be used for collaborative robots' intelligence grasping,but also make some contribution to related sub-domains,such as object recognition,pose estimation and path planning.The research results can increase the intelligence of collaborative robots to a certain extent,hence promote the extensive application of collaborative robots in industrial production.