Research On Imitation Learning Of Mobile Manipulators Based On Autonomous Learning Model

Mobile manipulators have well integrated mobility and manipulation capability.Thus mobile manipulators have great potential in lots of areas,such as components transportation,surface treatment of huge components and intelligent maintenance.But traditional planning motion for each task is not a technically and economically efficient way.With the development of robot imitation learning,non-technical robot users can improve the performance and capability of robot via demonstration according to the task requirements.As a result,robot applications can get rid of professional skills of users,which causing an improvement in both robot intelligence and efficiency,and bring robots close to non-technical people.However,current research on robot imitation learning mainly focuses on fixed manipulators.Fixed manipulators are dramatically different from mobile manipulators in the aspects of structure,tasks and working environments.Compared with imitation learning of fixed manipulators,imitation learning of mobile manipulators faces new challenges in autonomously learning demonstration trajectories and executing learned tasks.Imitation learning methods for fixed manipulators are not applicative to mobile manipulators,and there is still no imitation learning method for mobile manipulators.To handle these challenges,we aim to do research on imitation learning method of mobile manipulators from the aspects of unknown trajectories learning,trajectory tracking under constraints and imitation learning method for mobile manipulators,which would benefit the research fields like robotics and artificial intelligence as well as industrial and daily life.Trajectory learning is the key point for imitation learning of mobile manipulators.We research imitation learning models of both discrete and rhythmic trajectories based on Dynamical movement primitives(DMPs).By analyzing the process of imitation learning based on DMPs,we discuss two difficulties the trajectory learning models based on DMPs in terms of autonomy,that is,autonomous pattern judgment for unknown trajectories and parameters assignment for learning models.We solve the two problems via an autonomous learning model constructed by fusing Bayesian optimization and dynamical movement primitives.Finally,unknown trajectories are learned autonomously,and the effectiveness of the algorithm is verified via experiments.Task execution is an important part for imitation learning of mobile manipulators,and the key component of task execution for mobile manipulators is trajectory tracking.Mobile manipulators face inner and outer constraints during trajectory tracking,different constraints may cause different force to the robot.By minimizing the total constraint force,we obtain a proper priority assignment of different control primitives.An algorithm based on artificial potential field is proposed to construct jacobian matrices of different control primitives,and the task priorities are achieved via null space projection.Based on the above,a trajectory tracking algorithm based on multi-objective optimization and null space projection is constructed for mobile manipulators.Finally,the effectiveness of the proposed algorithm is verified by tracking a learned trajectory with a mobile manipulator under several constraints.The experiment results prove that the tracking algorithm is safe and the tracking error is 37.1% smaller than similar algorithms.According to the process of robot imitation learning,we analyze the difference between imitation learning of fixed and mobile manipulators based on their different in terms of structure,tasks and environment.On the basis of autonomous learning model and trajectory track algorithm,we propose an autonomous imitation learning method for mobile manipulators.With this method,several problems for imitation of mobile manipulators are solved,and the whole process of imitation leaning of mobile manipulators can be managed autonomously.Finally,according to the requirements of imitation learning of mobile manipulators,a wheeled mobile manipulator is design.Several experiments,such as clean whiteboard,pour water and pick and place,etc.,are carried out to verify the effectiveness of the propose algorithms and their applicability to different tasks.