Repetitive　motion　planning　in　semi-structured　environments　is　involved　in　many　robotics　applications,　such　as　manufacturing　and　in-orbit　service　of　spatial　robotic　arms.　In　these　scenarios,　the　environment　is　mainly　static,　but　it　may　contain　some　movable　obstacles　that　the　robot　needs　to　avoid,　and　the　planning　problems　that　the　robot　will　encounter　are　very　similar.　However,　most　current　sampling-based　motion　planning　methods　ignore　the　correlation　among　tasks,　and　instead　plan　each　task　instance　from　scratch,　or　assume　that　the　environment　is　highly　structured.　In　order　to　overcome　this　problem,　this　paper　proposes　an　experience-based　bidirectional　rapidly-exploring　random　trees*　(EBRRT*)　for　repetitive　path　planning　in　semi-structured　environments.　First,　an　adaptive　sampler　learned　from　the　demonstration　paths　by　a　Gaussian　mixture　model　(GMM)　is　constructed,　which　can　guide　the　planner　to　generate　samples　in　task-related　regions.　Second,　an　experience　graph　is　proposed　to　capture　the　exploration　of　configuration　space　by　previous　similar　tasks.　Subsequent　tasks　can　retrieve　locally　valuable　information　from　the　graph　to　avoid　exploring　the　configuration　space　repeatedly.　In　addition,　the　validity　of　edges　contained　in　the　experience　graph　is　lazily　verified　during　the　planning　process　so　that　the　experience　graph　can　adapt　to　small　environmental　changes.　Finally,　the　proposed　method　is　verified　in　a　2D　simulation　environment　and　a　7-DOF　manipulator　scene.　Experimental　results　show　that　the　proposed　method　can　return　a　high-quality　path　in　less　time　than　other　state-of-the-art　sample-based　motion　planning　for　repetitive　path　planning　in　semi-structured　environments.　©　2023,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG.