Reinforcement Learning And Intention Inference For Human-robot Interaction

The only way for intelligent system to complete complex task under dynamic uncertain environment is to have the functions of online learning and human-robot interaction. In the actual interaction process, to complete natural and effective interaction with human, it is necessary to correctly recognize and infer the user’s intention. How to construct such a theoretical architecture, which combines machine learning and human intention inference together for human-robot interaction tasks, is of great importance in academic and application.This thesis works on the background of intelligent robot working in the actual dynamic complex environment for interactive tasks, aiming at how to get the user’s intention effectively, so as to improve the convergence speed of the reinforcement learning algorithm. Based on the study of reinforcement learning and intention inference, a novel reinforcement learning and intention inference scheme based on human robot interaction is proposed. Experiments show the validity of the proposed methods. The main work is summarized as follows:1) A robot tracking control scheme was designed, which could adapt to the dynamic uncertain environmentSince the existing robot tracking control cannot adapt to the dynamic changing environment, a novel robot control scheme based on visual sensors was proposed. Combining particle filter and reinforcement learning algorithm, the scheme could deal with the complex dynamic uncertain environment.2) A grey prediction based intention inference algorithm was proposed to infer the walking intention of humans, in order to adapt to the change of movement.Without considering the walking style of humans, the existing control algorithms make certain hysteresis in the robot following task. An intention inference algorithm based on the theory of grey prediction model was designed to model human’s walking intention, which improved the robot following performance.3) A greedy guidance based fast interactive reinforcement learning algorithm was proposed to fully use human guidanceIn order to make full use of human guidance given in the process of interaction, so as to search more effectively in the state space and speed up the convergence, a greedy guidance based fast interactive reinforcement learning algorithm, which makes its application on real robot system into reality.4) An intention inference based interactive reinforcement learning algorithm was proposed, which could effectively identify humans’ real intention in the case of occasional wrong messagesIn the process of human-robot interaction, wrong information will be given by user due to various kinds of reasons. Considering this problem, this thesis designed two intention modeling methods, which included reward intention modeling and guidance intention modeling. Then combined with the previous proposed algorithm, an intention inference based interactive reinforcement learning algorithm was designed. Experiments showed that the algorithm could effectively recognize humans’ real intention even with the wrong information.