Swarm Deep Reinforcement Learning for Robotic Manipulation

Abstract

Deep reinforcement learning scheme, which combines both deep learning and reinforcement learning, enables robots to learn from exploration and flexibly performance in a range of different operational tasks under highly dynamic and complex environments encountered in daily life. However, robotic manipulation still face many serious threats due to inadequate data sharing between robots and concerns about data privacy and security. To privacy-protect the data of all owners, we propose a swarm reinforcement learning method, a decentralized deep reinforcement learning technology based on block chain. Specifically, each robotic agent controls the robot using actor-critic strategy optimization algorithm, and shares their learning experience (i.e. loss function gradient) through the blockchain network, and passes on a mature strategy model parameters to other agents. Experimental results indicate that our swarm reinforcement learning method can improve the learning process of several agents, and the more agents there are, the faster the learning speed will be.