Multi-agent Cooperative Control Under Constraints

With the rapid development of advanced communication technologies and intelligent computing technologies,multi-agent systems become a key factor to cross Artificial Intelligence 2.0,bringing new challenges and opportunities to frontier fields,such as communication networks,intelligent control,and bionic robots.Artificial Intelligence 2.0 points out that through the coordinated control of multiple agents,cluster-based collaborative intelligence beyond the single intelligence capability will emerge at the system level.In practical multi-agent missions,the agents are often required to meet velocity measurement constraints,output constraints,maneuvering constraints,etc.,which puts more demands on the design of control algorithms.For example,in some specific multi-agent tasks,the velocities and positions of mobile robots are required to stay within the limits of the specified constraints.Under this background,the concept of multi-agent control under constraints has emerged.This thesis aims to solve three important collaborative tasks:tracking,formation and circumnavigation with the considerations of various constraints on velocity measurements,output states and maneuvering configurations as follows:To solve the distributed tracking control problem under velocity measurement constraints,an adaptive high-gain observer is first proposed by using the equivalence and separation theorem to meet the velocity measurement constraint.Based on the proposed observer,a fully distributed tracking algorithm is designed via state and output feedback under directed communication networks.In order to compensate the uncertainty of the system model,the controller combines the adaptive neural networks to obtain the approximation ability,and finally the semi-globally uniform ultimate boundedness of the closed-loop errors is proved.Focusing on the distributed tracking control problem under system output constraints,a distributed finite-time sliding-mode estimator is designed under the condition that the dynamic leader information can only be acquired to a part of the followers,which can accurately obtain the desired trajectory for each agent.In order to meet the output constraints,an integral barrier Lyapunov function for distributed systems is designed,and a distributed tracking control algorithm is proposed.Note that most existing research on barrier Lyapunov function is only for one single system,and because of the uniqueness of its design process,this function cannot be directly applied to distributed systems.Theproposed algorithm extends the results in the literature by combining the structure of distributed observers and controllers,which also relieves the conservatism of the traditional barrier Lyapunov function.In the existing formation control results,most formation configurations are single,static and difficult to reconstruct.In this approach,a new two-layer formation structure is proposed and defined,which can characterize the complex multi-agent heterogeneous models and group behaviors,giving a unified description of a variety of basic distributed issues.Based on this two-layer structure,the multi-agent distributed constant and timevarying formation control problem with model uncertainties under velocity measurement constraints is solved.Due to the natural characteristics of the two-layer structure,the proposed distributed output feedback algorithm can be applied to the formation tasks of heterogeneous agents.To articulate the distributed circumnavigation control under the constraints of maneuvering configuration,the trajectory analysis,task planning and cooperative circumnavigation control design in three-dimensional space are studied.The proposed control law enables the networked agent systems to fly around the target while maintaining the geometric layout of the airway queue.Further,the proposed three-dimensional circumnavigation control algorithm is applicable to dynamic targets and solid elliptical orbits.Near-zero fuel consumption can be achieved for the satellite missions,and the closedloop system errors converge to the origin exponentially.