The Research On Grasp Planning By Convolutional Neural Networks

At present,industrial robots are widely used in vehicle manufacturing,food processing,electronics,medicine and other industries due to their highly automated and flexible characteristics.Compared with the traditional manual,robots have better stability and higher working efficiency.Compared with the non-standard equipment in the traditional automated production line,robots have lower cost,stronger versatility and better flexibility.However,it must be recognized that the current technological development level of industrial robots cannot meet the requirements of national and enterprise intelligent manufacturing transformation.Investigating its root cause,the biggest bottleneck of the robots is not that its accuracy,speed or stability,but the robot's ability to interact with the physical world,that is,the robot's perception ability,analysis ability and decision-making ability are not enough to handle complex unstructured application scenarios.In view of this,this research takes the robot grasp planning problem in the unstructured environment as content,combined with the thought of machine learning,using the convolutional neural network algorithm architecture to build a robot grasping planning solution,trying to improve the ability of robots to grasp objects known and unknown and simplify the using of the robots.Through research on the relevant content,this research divides the grasp planning problem into two stages: grasp quality evaluation and grasp pose candidate generation.Grasp quality evaluation analyzes the set of candidate grasp gestures and input depth images through a convolutional neural network to generate pseudo-confidence rankings,thereby pick grasp gesture with the highest success rate.Grasp pose candidate generation algorithm analyzes the input image features or grasp pose evaluation results to sample the complete grasp space,and generates a high-quality small-scale sample set to reduce the number of grasp quality evaluations and improve the efficiency of the entire solution.Due to the difference between the large-scale general grasp training dataset used to train the original model and the current grasp scenario,it may cause the network model to fail to achieve the expected results when processing the actual grasp scenario.By building an automated training sample collection algorithm,the original network model trained by the universal training dataset can perform transfer learning for a specific application scenario,which can improve the prediction accuracy of the network for this scenario.For the verification of the grasping planning algorithm,this research built a simulation experiment platform based on the Gazebo environment.The experiment of the grasping planning algorithm was carried out for known and unknown objects.Experimental results show that the algorithm described in this paper is effective and efficient,and the grasping process is stable and dexterous.