Autonomous learning of robot pushing and grasping skills in the cluttered environment has been widely studied. The cooperation between them is the key to improving grasping efficiency. In this article, a deep reinforcement learning algorithm GARLDQN based on the generative adversarial network and model generalization is proposed. Firstly, the generated adversarial network is embedded into the traditional DQN to train the coevolution between pushing and grasping. Secondly, some parameters in MDP are formulated based on the goal object, and the hindsight experience replay mechanism (HER) is used for reference to improve the sample utilization of the experience pool. Then, according to the image state, a random (convolution) neural network is introduced to improve the generalization ability of the algorithm. Finally, 12 test cases are designed and compared with the other four methods in terms of grasp success rate and average motion times. In the regular block cases, two indicators are 91. 5% and 3. 406, respectively. In the daily tool scene, two indicators are 85. 2% and 8. 6, respectively. These results show the effectiveness of the GARL-DQN algorithm in solving the problems of robot pushing and grasping cooperation and model generalization.