This　paper　focuses　on　the　learning　action　selection　in　behavior-based　autonomous　mobile　robot.　Autonomous　mobile　robot　needs　a　large　space　to　store　the　state-action　pair　in　the　application　of　tabular　Q-learning.　Neural　network　has　a　good　ability　of　generalization,　so　in　this　paper　Q-learning　based　on　neural　network　is　developed　which　has　a　good　ability　to　approximate　to　Q-function.　The　Q-learning　based　on　neural　network　is　applied　to　autonomous　mobile　robot　for　goal　directed　obstacle　avoidance.　Results　of　simulation　show　that　the　mobile　robot　can　learn　to　select　proper　actions　itself　to　accomplish　the　task　autonomously.