Research On Distributed Cluster Control With Obstacle Avoiding Technology Based On Flocking Algorithm

Distributed multi-agent group motion control technology is the technique basis of a series of future application scenarios,such as large-scale drone cluster search,Large-scale autonomous formation performance technology,ultra-long-distance drone cluster relay communication and so on.The design of building a multi-agent group equipped with obstacle avoiding ability is almost the study focus nowadays.The existing research was based on modeling the organism cluster motion,transforming obstacle into a virtual agent to realize cluster motion control.This method was called Flocking algorithm.However,this algorithm was only suitable for convex type obstacles and some kind of non-convex obstalces.As for the narrow obstacles which could almost allow a few agents passing through,this algorithm is not working.Then,the existing technique could not combine the different types of obstacle avoiding method together.So many research comes to deep learning for solution.However,it does not work well.Focusing on narrow obstacle avoiding problem and Flocking based on deep learning algorithm is the main contribution of this paper.Firstly,the narrow obstacle passing problem is a theoretical flaw in the existing Flocking algorithm.Therefore,in a theoretical sense,this scheme can fully improve the existing Flocking algorithm.Then,the existing formula method requires different obstacle avoidance solutions according to different obstacles,which is very complicated.If the agent can adapt to different obstacle types through machine learning,and learn this obstacle avoidance solution,it can greatly reduce the burden of the agent on obstacle recognition,which is more conducive to deployment and implementation in a practical sense..Through analysis,the existing Flocking is controlled by maximize the self benefit not the nearby group benefits,so resulting in the dilemma of the existing Flocking in narrow obstacles situation.Through analysis and proof by game theory in this paper,when facing this kind of obstacles,each agent among the group should follow the rule that the closest agent towards target should move first.Simulation of different number of agents and different dimensions shows the effectiveness.Aiming at the problem of narrow obstacles,this paper proposes a Flockingmotion control scheme based on game theory.In order to combine the different types of obtacles avoiding methods together,this paper proposes a two-stage deep learning Flocking algorithm.This paper finds that the existing deep learning Flocking algorithm could not describing the real surroundings of agents.So,this paper proposed a feature engineering based on the motion direction.It divides the agent motion directions into different types,and through deep learning to classify the correct directions which is the first stage,and then using another deep learning network to regress the length of acceleration in the classified direction that is the last stage.After a series of simulation,this algorithm realizes a Flocking with obstacle avoiding motion control technique in less complexity.