In　this　paper,　we　develop　an　improved　data-based　integral　policy　iteration　method　to　address　the　robust　control　issue　for　nonlinear　systems.　Combining　multi-step　neural　networks　with　pre-training,　the　condition　of　selecting　the　initial　admissible　control　policy　is　relaxed　even　though　the　information　of　system　dynamics　is　unknown.　Based　on　adaptive　critic　learning,　the　established　algorithm　is　conducted　to　attain　the　optimal　controller.　Then,　the　robust　control　strategy　is　derived　by　adding　the　feedback　gain.　Furthermore,　the　computing　error　is　considered　during　the　process　of　implementing　matrix　inverse　operation.　Finally,　two　examples　are　presented　to　verify　the　effectiveness　of　the　constructed　algorithm.