Game-theoretic robotic offloading via multi-agent learning for agricultural applications in heterogeneous networks

Abstract

Intelligent robotics, as a frontier field of widespread attention, is increasingly applied to realize cyber-physical-social system (CPSS). Taking Agriculture 4.0 as one use case, using various agricultural robot applications can reduce labor costs and improve operation efficiency, but limited by the onboard resources, these mobile robots need to make full use of network and computation resources provided by infrastructures in the smart farm to enhance their capabilities.On the one hand, the heterogeneity of networks brings more options for access selection; however, on the other hand, the mobility of robots and the Quality of Service (QoS) requirements of different agricultural applications pose great challenges to the current offloading work. Therefore in this paper, we introduce a stochastic game framework to address these challenges faced by smart agriculture, and propose a multi-agent task offloading and network selection (MATONS) scheme, which utilizes reinforcement learning to offload tasks from robots to central cloud or edge servers. The robots cooperatively optimize the QoS perceived and the energy usage by adopting a joint Nash equilibrium strategy, which requires the decision of offloading along with the optimal transmission technology and corresponding access nodes, i.e., WiFi or cellular network. Simulation results verify that MATONS has great convergence and stability under different scales of robots in the farm, and can achieve superior performance than mainstream algorithms in various scenarios.