Understanding3Dscenesisacornerstoneofautonomousmobilerobotnavigationandinteraction, especiallywithincomplexindustrialenvironments.Traditionaloccupancy-basedSLAMmethods typicallyemphasizespatialgeometry,yetoftenlacksemanticcontext—anessentialcomponentfor informeddecision-making,object-levelreasoning,andprecisetaskexecution.Toovercomethis, varioussceneunderstandingapproacheshavebeenproposed;however,theyareoftenconstrainedby their relianceonlargeannotateddatasets,challengesindetectingcustomornovelobjects,di!culties in sensorfusionduetonon-overlappingfieldofviews,andconstraintswithreal-timeperformance. These shortcomingssignificantlylimittheadaptabilityandscalabilityofsceneunderstandingin real-worldapplications.Thisthesispresentsanovelsystemthatequipsmobilerobotswiththe ability togeneratesemanticallyrich2Doccupancymapsenhancedby3Dobjectreconstructions through amulti-viewfusionframework.Crucially,itenablesthedetectionofnovelobjectswith minimal supervision,eliminatingtheneedforlargeannotateddatasets,humanannotationsand reducing deploymentoverhead.Byaggregatingobjectdetectionsfrommultipleperspectives,the systemdeliversmoreaccurateandresilient3Dreconstructions.Enhancing2Doccupancymaps with 3Dobjectreconstructionsintroducessemanticstructureandobject-levelawarenessintothe spatial map,enablingricherunderstandingofthescene.Theproposedsystemrepresentsastep towardsreal-time,andscalable3Dsceneunderstanding—pavingthewayformoreintelligent, autonomous, andadaptablemobilerobotsindynamicindustrialsettings.