As　for　autonomous　driving　in　urban　environments,　it　is　of　significance　to　accurately　capture　the　position　and　pose　of　vehicles.　Those　information　can　assist　self-driving　system　in　making　right　decisions　to　avoid　potential　risks.　Currently,　3D　point　clouds　captured　by　laser　scanners　are　widely　used　in　self-driving　systems　to　sense　the　real　environment.　Therefore,　in　this　paper,　we　propose　a　deep　reinforcement　learning　framework　for　vehicle　detection　and　pose　estimation　by　using　3D　point　clouds.　Specifically,　to　estimate　the　pose　of　vehicles,　we　propose　to　design　a　rotation　action　in　Deep　Q　Network　(DQN).　In　addition,　by　considering　the　whole　detection　procedure　as　Markov　Decision　Process　(MDP),　our　intermediate　detected　results　can　further　improve　the　detection　performance　of　our　proposed　method.　The　evaluations　are　carried　on　outdoor　point　cloud　scenes　captured　by　VMX450　laser　scanning　system.　The　experimental　results　demonstrate　the　satisfied　performance　on　vehicle　detection　and　pose　estimation.　©　2020,　Springer　Nature　Switzerland　AG.