Grasp Control Method for Robotic Manipulator Based on Federated Reinforcement Learning

摘要

With development of mobile edge computing (MEC) and the Artificial Intelligent Internet of Things (AIoT), how to use intelligent devices to better apply to industrial scenarios has become a critical factor in promoting industry to reduce costs, increase efficiency, and enable development. In this paper, we analysis the problem of robotic manipulators working in dynamic environments and propose a model based on federated reinforcement learning (FRL) to control robotic manipulator grasping. In the proposed FRL framework, the optimal control policy is learned independently by multiple agents on their respective AIoT devices. Also, we use federated learning to ensure the safety of data and improve the reward function to accelerate the learning of model. We evaluate the performance of the proposed method and compare it with the traditional method based on RL. The experimental results show that proposed algorithm significantly improves the model training convergence speed by about 64% and the task average reward performance index by about 36% compared with baseline.