In　this　paper,　the　model-free　robust　control　problem　is　investigated　for　nonlinear　systems　with　a　relaxed　condition　of　initial　admissible　control.　An　advanced　integral　reinforcement　learning　method　is　developed,　which　merges　the　adaptive　network-based　fuzzy　inference　system　(ANFIS)　and　pre-training　of　the　initial　weights.　To　loose　the　condition　for　choosing　the　initial　control　law,　pre-training　of　initial　weights　is　established　by　utilizing　the　ANFIS　to　provide　the　information　corresponding　to　the　system　model,　which　is　applicable　to　the　model-free　issue.　Based　on　the　actor-critic　structure,　the　approximate　optimal　control　law　is　obtained　by　employing　adaptive　dynamic　programming.　Redesigning　the　obtained　control　law,　the　robust　controller　can　be　derived　to　stabilize　the　system　with　the　uncertain　term.　Eventually,　two　examples　are　utilized　to　verify　the　effectiveness　of　the　constructed　algorithm.