The　performance　of　unmanned　ground　vehicle　(UGV)　formation　is　crucial　for　large-scale　material　transport.　In　a　non-communicative　environment,　visual　perception　plays　a　central　role　in　formation　control.　However,　due　to　unstable　lighting　conditions,　dust,　fog,　and　visual　occlusions,　developing　a　high-precision　visual　formation　control　technology　that　does　not　rely　on　external　markers　remains　a　significant　challenge　in　UGVs.　This　study　developed　a　new　UGV　formation　controller　that　relies　solely　on　onboard　visual　sensors　and　proposed　a　teacher-student　training　method,　TSTMIPI,　combining　the　PPO　algorithm　with　imitation　learning,　which　significantly　improves　the　control　precision　and　convergence　speed　of　the　vision-based　reinforcement　learning　formation　controller.　To　further　enhance　formation　control　stability,　we　constructed　a　belief　state　encoder　(BSE)　based　on　convolutional　neural　networks,　which　effectively　integrates　visual　perception　and　proprioceptive　information.　Simulation　results　show　that　the　control　strategy　combining　TSTMIPI　and　BSE　not　only　eliminates　the　reliance　on　external　markers　but　also　significantly　improves　control　precision　under　different　noise　levels　and　visual　occlusion　conditions,　surpassing　existing　visual　formation　control　methods　in　maintaining　the　desired　distance　and　angular　precision.