Research On Key Technologies In Cloud Robotic Lifelong Learning System Based On Federated Learning

Cloud robotic is a new cross research field in recent years.Cloud robotic is a kind of technology that helps to complete many complex tasks on local robots with the help of cloud computing resources.One of the vision of cloud robotic system is to build cloud brain and realize lifelong learning of robots.To achieve this,this paper proposes learning frameworks and key algorithms of robotic lifelong learning system based on Federated learning.The simultaneous interpreting of di erent dimension data(homologous isomerism),di erent meaning sensing data and di erent dimension data(heterogenous isomerism)of the same sensor are analyzed,corresponding frameworks and algorithms are proposed.Finally,the construction of cloud brain in cloud robot system is realized.With the shared model of cloud,the cloud robotic system can realize the fusion,evolution and migration of knowledge,and e ectively improve the learning e ciency of local robots.This paper mainly completed the following work:1.On the premise of data homology and isomorphism,we propose lifelong federated reinforcement learning(LFRL)for cloud robotic lifelong learning.In addition,The cloud knowledge fusion algorithm is presented and two available knowledge migration methods are provided.2.In view of the situation of data homology and heterogeneity,we present the heterogeneous federated learning(HFL)for cloud robotics to realized lifelong learning.Among this,cloud knowledge fusion algorithm and knowledge migration algorithm are proposed.3.In view of the heterogeneous data in the lifelong learning system of cloud robotics,this paper proposes the federal imitation learning(FIL).Moreover,the work proposed the cloud knowledge fusion algorithm in FIL and provides the available knowledge transfer methods.