Research On Collaborative Task Offloading Strategy Of Cloud Robot Under Multi-objective Constraints

With the rapid development of technologies such as electronics,artificial intelligence,and the Internet of Things,industrial scenarios are experiencing the fourth revolution led by smart manufacturing currently,which is prompting the transformation of traditional factories into smart factories.robots can independently perform small-scale data processing.In order to solve the limitation that the limited resources of robots are not enough to support intelligent decision-making and real-time response,the concept of multi-robot system is proposed;Cloud and edge computing are successively extended to the system,forming a cloud-robotics-oriented end-edge-cloud collaborative computing architecture.This architecture realizes the three-layer resources efficiently and collaboratively providing services,and can effectively exert the maximum efficiency of each resource,thereby contributing to the digital and intelligent transformation of the industry.However,the heterogeneity of each resource and the dynamic diverse task requirements make the study of task offloading more complicated.This paper mainly studies the task offloading problem of cloud-robotics-oriented end-edge-cloud collaborative scenarios in the Industrial Internet.The specific work is as follows:Combined with the actual industrial logistics scenarios,this paper studies a cloud-robotics-oriented end-edge-cloud collaborative task offloading architecture,which supports comprehensive collaborative offloading of multiple resources,allowing the main robot to offload tasks to auxiliary robots,edge and cloud.Aiming at the task offloading problem under this architecture,with the goal of minimizing task completion delay and task completion energy consumption,the offloading problem is modeled as a multi-objective optimization problem.In this paper,We based on Genetic Algorithm(GA)propose a heuristic collaborative offloading algorithm ASA＿GA to solve the multi-objective task offloading problem.In order to jump out of the local optimal solution and speed up algorithm convergence,this paper improves the algorithm from two aspects of the new solution acceptance rule and the adaptive crossover mutation.Compared with the same type of algorithms,the improved ASA＿GA algorithm has certain advantages in both optimization effect and algorithm running time.Considering the powerful learning strategy capability of deep reinforcement learning algorithms,this paper based on Deep Deterministic Policy Gradient(DDPG)propose a deep reinforcement learning collaborative offloading algorithm NPE＿DDPG.The algorithm supports offline training and then online decision-making,transforming the complexity of online decision-making into the complexity of offline training,so that it can quickly respond to changes in resource and task requirements in the Industrial Internet.In this paper,we improve the algorithm from the perspective of enhancing training stability and accelerating convergence,and integrate the proposed Critic Update Smoothing and Prioritized Experience Replay into DDPG.Compared with the same type of algorithms,NPE＿DDPG has more stable convergence effect and better optimization effect;Compared with the ASA＿GA algorithm,NPE＿DDPG has a similar optimization effect but greatly reduces the running time of the algorithm.