To address the problem of traditional navigation about unmanned vehicles in narrow turning work environments, such as the inability to construct maps or the construction of maps with excessively large obstacle expansible radii, as well as errors in positioning and control, resulting in collisions with obstacles and ineffective completion of navigation tasks, a method combining the A* algorithm and deep reinforcement learning is proposed. The path generated by the A* algorithm is discretized, and periodically selected path points are used as target points for the deep reinforcement learning algorithm. A subgoal-driven DQN algorithm is designed, and on this basic a neural network is established. The narrow turning environment is constructed using Gazebo software, and the unmanned vehicle is trained using the subgoal-driven DQN algorithm, DQN algorithm without subgoals, DDPG algorithm, and SAC algorithm. By comparing the convergence speed, execution steps, and navigation success rate, it is demonstrated that the subgoal-driven DQN algorithm performs best in completing the navigation task in narrow turning environments. The training results of the subgoal-driven DQN algorithm are transferred to a new test scenario with smaller space and more turns, and the test verifies that the unmanned vehicle can successfully complete the navigation task, proving the high scalability of the subgoal-driven DQN algorithm.