An　intelligent　control　architecture　with　reinforcement　learning　was　designed　based　on　a　behavior-based　architecture　to　improve　the　learning　ability　of　mobile　robots.　Normal　tabular　Q-learning　can　only　be　applied　to　discrete　states　and　requires　a　large　memory.　Since　neural　networks　have　good　generalization,　a　Q-learning　system　was　developed　based　on　a　neural　network　for　obstacle　avoidance　of　mobile　robots.　Experiments　show　that　the　mobile　robot　can　then　learn　to　avoid　obstacles.