The invention discloses a mechanical arm pushing and grabbing system suitable for a dense environment. The system comprises a PC, a color depth camera, a mechanical arm and an intelligent clamping device; the software end comprises a CoppeliaSim simulation platform; the color depth camera is connected with the computer through a USB, and the mechanical arm is connected with the PC through a local area network; the color depth camera acquires a color RGB image and a depth image; the computer executes a mechanical arm control program; the mechanical arm executes actions; the intelligent clamper is used for grabbing or pushing the object block; and the CoppeliaSim simulation platform comprises a simulation module, a calibration module, an image preprocessing module, a feature extraction module, a decision network module, an action strategy module, a robot I/O module and a robot module. According to the method, simulation and reality are combined, robot damage caused by training is reduced, and the grabbing speed is increased; and the DQN of deep reinforcement learning is utilized, action semantics are dynamically planned in real time according to the current object environment, pushing and grabbing are combined, meanwhile, the most suitable grabbing direction is planned, and the success rate of grabbing the complex environment by the mechanical arm is greatly increased.