Research On Mobile Robot Control Methods Based On Learning Human Strategy

With technological changes such as artificial intelligence and intelligent system,the control method of mobile robots has become a research hotspot in the frontier fields of control science and engineering.Mobile robots have a wide range of application scenarios and huge development potential.However,most mobile robots currently work in structured scenarios,completing relatively simple repetitive tasks based on pre-programming by technicians.For example,inspection robots generally patrol according to a fixed path,and food delivery robots are often only suitable for specific restaurant environments.Therefore,how to effectively combine the scene perception,path planning and motion control technologies of mobile robots and design controller based on learning human strategies is a very valuable research direction.In this process,users do not need professional controller design knowledge and parameter adjustment skills,which helps to further reduce the use threshold and product cycle of mobile robots.Based on the above research background,this paper mainly studies how to teach mobile robots to learn the control strategies of human experts and apply them to basic control tasks,including path following control,and fast and stable control of statically unstable and dynamically stable mobile robot platforms(such as mobile balance car)to effectively improve the robustness and generalization ability of mobile robots in unstructured environments.This paper studies mobile robot control methods based on learning human strategies from three aspects: motion control methods for two-wheel differential mobile robots,model predictive control methods for unmanned logistics vehicles,and learning from demonstration for mobile balancing vehicles.(1)The motion control method for a two-wheel differential mobile robot.This part can be divided into point stabilization control problem and path following control problem.For point stabilization control problem,we designed a controller and conducted simulation experiments to verify it.For path following control problem,we designed a controller to follow the path from point A to point B.The path-following experiment,the experiment of bow-shaped full-coverage trajectory following,and obstacle circumvention experiments verify the effectiveness of our proposed method.(2)Model predictive control method for unmanned logistics vehicles.First,we established the kinematics model of the unmanned logistics vehicle.Second,we designed an improved model predictive controller through the steps of deriving the predictive equation,optimizing the solution and adding the feedback mechanism,and conducted simulation experiments to verify it.Finally,based on the unmanned logistics vehicle platform of our research group,we verified the effectiveness of the proposed control method.(3)Learning from demonstration control method for mobile balance car.For some nonlinear systems that are difficult to build a complete and accurate model,the control strategy based on learning from demonstration is a novel and effective control method.Human experts only need to perform multiple demonstrations,and the controller parameters based on learning will be automatically calculated.In this process,users neither need professional controller design knowledge,nor controller programming and parameter adjustment skills.Therefore,we propose a method based on learning from demonstration that can realize the fast and stable control of statically unstable and dynamically stable mobile robots(such as mobile balancing vehicles).In order to verify the correctness of the theoretical analysis,we applied this kind of controller designed based on learning from demonstration to the control task of a mobile balancing car,and verified its effectiveness through experiments.Compared with the traditional control method,this method has a certain generalization ability,and can promote the situation that did not happen in the demonstration process.In summary,this paper focuses on control methods based on learning human strategies for mobile robots.It is essentially a method that does not require complex modeling and only needs demonstrations,and can be extended and applied to other mobile robot platforms.The most important thing is that we summarize the reasons for the good or bad learning results based on the learning of human policy controllers.The case of poor learning may be that the dynamic model has not learned well,or it may be that the control strategy has not been learned well.Moreover,the constraints on the convergence conditions are very important for learning results.We qualitatively analyzed how to learn better to adapt to different control platforms and improve the generalization ability of the controller.