Research On Intelligent Perception And Autonomous Control Of Manipulation

Intelligent robot technology is a cross-disciplinary field.The technology of environment recognition and autonomous learning is the key to realize robot autonomous control and improve robot intelligence.Through the three-dimensional environment modeling technology and the target recognition technology,the robot can quickly perceive the environment information and get the original data for decision.Through the machine learning,state reasoning and other data analysis and forecasting technology,the robot can update its own action strategy according to the environment information and its own state,and carry out autonomous planning and control on the next action.In this paper,we design and implement the design and implementation of the architecture of the robot system,the perception and modeling of the environment,the recognition of the target object and the learning of the robot action while the classical robot environment sensing technology and the robot autonomous learning technology are absorbed.Based on the distance data,the dynamic environment perception and modeling algorithm is implemented,and the autonomous action learning of the robot is studied,and the following aspects are mainly studied:1.According to the functional requirements of intelligent robots,the architecture of robot system which can carry out environment-aware and action learning is designed and realized by using open-source hardware and software.2.Based on the characteristics of object manipulating robot,an environment-aware algorithm based on distance sensor is designed.In the aspect of data acquisition,multi-dimension data acquisition improves the blind area in traditional unidirectional data acquisition process.To solve feature extraction problem,a hierarchical plane extraction algorithm based on Hough transform is proposed,which can be carried out synchronously in the sensing process,and the efficiency of feature extraction is improved.In the aspect of 3D object recognition,the method of depth projection reduces the dimension of training data and improves the operation efficiency by mapping the three-dimensional data to the two-dimensional plane with less loss of accuracy.3.Based on MDP and reinforcement learning framework,a robot action learning algorithm based on environment modeling and action abstraction is proposed to improve the versatility and anti-jamming ability of the robot action learning model,And reduces the computational overhead.A training set construction method based on data flow model is proposed,which can eliminate the junk data in the training set in time.The multi-model voting algorithm is improved,and its ability to predict continuous data sets is improved.4.The hardware and software of the manipulator are designed and realized in thispaper.The manipulation has the software and hardware foundation to realize the object manipulation,the environment perception and the autonomous control.environment perception and autonomous control.With the cloud host to help run the algorithm,and with some help of artificial support,the robot realized the autonomous learning and control that based on the environmental model.