Research On Precise Learning From Demonstrations For Space Manipulators

In recent years,with the deepening of space exploration and robot technology,the tasks of space manipulators are being performed from the traditional capture tasks to precise operation tasks,such as on-orbit refueling,on-orbit assembly,and on-orbit maintenance.The increasing tasks bring a more significant challenge of the operation methods of space manipulators on adaptability,ease to operate,and precision.A precise Learning from Demonstrations(Lf D)method is proposed for precise operation tasks of space manipulators.Compared with traditional Lf D methods,the three steps of the process are improved: perceiving the teaching movement,building the skill models,and reproducing the task skills.The method is verified by the simulation and experiment.Firstly,a precise Lf D system for space manipulators was designed.For the onorbit maintenance task,the precise Lf D system is divided into the precise Lf D module and the on-orbit teleoperation supervision module.The precise Lf D module is divided into the ground part and the on-orbit part.The ground part is used to perceive the teaching movement of task experts and build skill models;the on-orbit part is used to reproduce the task skills.Then,for task experts’ precision,comfort,and safety during operation,a hybrid motion mapping method for the manipulator is introduced.In the method,a convex hull-based precise motion mapping method is used in the free workspace of the manipulator to improve the mapping accuracy;a virtual impedance force feedback mapping method is used in the workspace near obstacles for operation safety.The on-orbit teleoperation supervision module is used to supervise reproducing task skills and deal with emergencies.Using the models generated by traditional skill modeling methods cannot estimate whether the space near the reproduced trajectories is safe.To solve the problem,a skill modeling strategy based on the task safety space estimation is proposed.The main part of the method is the skill modeling method based on the Gaussian process.Using the Gaussian process skill model and the state of the experiment,a stochastic desired trajectory can be reproduced.The variance part of the trajectory can be used to estimate the task safety space,which is the foundation of the skill optimal control method.In addition,the skill modeling method based on velocity field can improve the stability and convergence speed near the target position.The skill model updating method based on the dynamic system can design modulation factors based on the obstacles for the existing skill models,and reduce repeated teaching.To solve the problems of joint desired torque fluctuation,excessive energy consumption,and poor precision caused by dynamics uncertainty while reproducing the task skills,a precise skill optimal control method based on dynamics uncertainty is introduced.Firstly,an optimal trajectory generation method based on the task safety space estimation is proposed to optimize the desired trajectories,reduce the joint expected torque fluctuation and energy consumption.Secondly,for the problem of insufficient precision caused by the dynamics uncertainty,the stochastic dynamics model is established with the data of actual manipulators,and a stochastic optimal control method is proposed.The method can overcome the dynamics uncertainty disturbance and improve the trajectory tracking accuracy.Finally,the effectiveness of the above method is verified by simulation.To verify the precise Lf D method for space manipulators,a ground experiment of locating pre-screwed bolts and pre-grasping with obstacle avoidance is designed in the thesis.In the experiment,the dynamics parameters of the manipulator have been changed by grasping a hand drill.The manipulator can still reproduce the task skills using the precise Lf D method,and the precision can meet the task requirements.Moreover,the method can significantly reduce the joint desired torque fluctuation and energy consumption.Finally,the effects of different stochastic desired trajectories and sensor noise on the stochastic optimal control method based on dynamic uncertainty are analyzed by additional experiments.