RLNN: A force perception algorithm using reinforcement learning

Force perception is one of the important research branches in human-computer interaction and compliant control of contact-rich robots. Force perception without robot end-effector sensors has attracted increasing attention in recent years, but the existing research methods do not consider the feature selection of variables in the process, and redundant dimension will increase the force perception cost of robots. To solve the above problems, we propose a new algorithm framework of Reinforcement Learning Neural Network (RLNN), which can realize the force perception of contact-rich robot. And it has the advantages of dimensionality optimization of input variables and lightweight network structure. Feature selection experiment, network structure experiment and different variable prediction experiment are conducted respectively, which proves the feasibility of our proposed algorithm framework. For force perception or prediction in robots, our research method and experimental results provide certain significance on how to balance perception dimension and perception cost.