Move-stop motion planning is critical to boosting performance of multi-arm picking robots in greenhouses where tomato plants are densely grown. This paper proposed a deep reinforcement learning-based vehicle motion decision model for a double-arm picking robot, with the goal to minimize stopping frequency and missed objectives while optimizing double-arm task allocation. The model initially generates a 2D projection map of the objective fruits on both crop rows, followed by a sequence of parking nodes for the vehicle using the DDPG algorithm. We examined four reinforcement learning models in an autonomous greenhouse simulation environment, and the convergence performance of DDPG outperformed the TD3, SAC, and PPO models by 15.24%, 27.93%, and 40.97%, respectively. The findings indicated that the parking planning approach put forward in this research can significantly raise the operational efficiency of the double-arm picking robot.