Research On Learning From Demonstrations And Intelligent Control Methods For Robotic Manipulation

Robot manipulation technology has been widely applied in different scenarios,which is a hot research topic with enormous potential.But,nowadays,in most applications,robots are deployed and restricted to a structured environment to perform some single,repetitive,and regular tasks using the pre-programming strategy.Integrating artificial intelligence into robot technology to realize the understanding and learning of the robot for operation skills without tedious manual programming and debugging is an important way to improve the intelligence level of robots,make the robots simple and fast to utilize,and promote industrial upgrading.At present,with the rapid development of artificial intelligence technology,robot operation learning technology has become a research hotspot in the frontier field of robotics.Based on the above research background,this dissertation mainly studies on how to make robots understand and learn the basic motions in task manipulations,including point-to-point motions,trajectory tracking motions,compliant motions,and hand-eye coordination motions.This thesis effectively solves the trade-offs between the speed,the accuracy,and the robustness of point-to-point motion learning,the generalization problem of trajectory tracking motion learning,the compliant motion learning,and the problem of incremental updating model parameters.In general,the main research contents and contributions of this dissertation are as follows:(1)A point-to-point motion learning framework based on a feedforward neural network is proposed.This method uses the extreme learning machine(ELM)to model and learn human demonstrated actions,and integrates Lyapunov's stability criterion into the constraints of the neural network optimization process to ensure that the pointto-point motions are stable and can converge to the desired position.The proposed method can not only improve the speed of learning but also show better robustness and accuracy in reproduction.Compared with other methods,it has a simple structure and high calculation efficiency.(2)After implementing the point-to-point learning control,this thesis extends the learning strategy to the problem of accurate tracking of complex trajectories further.In response to this problem,the strategy of the convergence for the position error is directly learned using the trajectory tracking demonstration data,and through learning,the obtained high-level robot speed controller that can make the actual trajectory converge to the ideal trajectory in real-time;then we use Lyapunov theory to derive the stability constraints of the related parameters in the control algorithm,and prove that the designed speed control law can ensure that the position error and the speed error both converge to zeros in the end.Because of the learning of error convergence,this method shows a good learning and generalization ability,that is,after learning a tracking strategy of a fixed shape trajectory,it can be applied to track other desired trajectories without any controller adjustments.(3)The compliant motion contains not only the motion trajectory but also the compliance information of force perception.This dissertation proposes a sequential learning control framework that combines motion trajectory and compliant force behavior.The learning framework collects data such as forces,displacements,speeds,and accelerations demonstrated by the teacher through a low-cost teaching interface,and then we design a sequential neural network to encode the robot's motion and propose a variable impedance learning algorithm based on the teaching force to estimate the variable damping and stiffness matrices in three directions.Both the motion trajectory and the compliance behavior learning adopt an online update strategy in a recursive form.Compared with the traditional strategy,which first collects data then trains offline,this update method not only improves the efficiency of teaching and learning but also avoids excessive or insufficient offline data collection questions.(4)A hand-eye coordination motion learning framework is proposed.The framework can automatically learn the image-based hand-eye coordination grasping operation based on demonstration data without prior hand-eye calibration.By adopting the graspable position learning in the image plane and the intelligent function approximator,the desired pixel position in the traditional servo needs not to be manually set and the hand-eye coordination needs not to be calibrated in advance.In addition,by designing an adaptive parameter adjustment strategy based on fuzzy logic rules,the problem of different convergence speeds of the data-driven controllers caused by different teaching data is overcome.Thus,the convergence speed of hand-eye coordinated motion is effectively improved.