We　present　a　framework　for　generating　gait　actions　of　a　virtual　character　in　real-time.　Integrate　physical　IK　into　the　model　make　the　neural　network　learn　the　movements　of　the　main　joints　more　specifically.　Our　method　takes　the　action　state　of　the　character＇s　previous　frame　and　the　terrain　of　the　scene　as　input　user　controls　and　automatically　generates　high-quality　motions　to　achieve　the　required　user　controls.　The　entire　network　is　trained　end-to-end　on　a　large　data　set　consisting　of　gait　movements.　Therefore,　our　method　can　automatically　generate　characters　to　adapt　to　the　movement　of　different　terrains.　Compare　with　the　Gaussian　process　and　EDR　model,　our　network　structure　can　produce　higher　quality　results.　Our　work　is　best　suited　for　controlling　characters　in　interactive　scenes,　such　as　computer　games　and　virtual　reality　methods.　©　2021,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.