A vision inspection robot online trajectory planning algorithm based on generative adversarial neural networks was proposed for industrial robots to address the problems of low trajectory planning efficiency and poor adaptability of planning results in unstructured environments. Firstly, a point cloud dataset construction method based on robot operating system simulation was proposed. Secondly, a new coding-decoding structure-based generative adversarial networks was proposed to generate robot inspection trajectories end-to-end using the point cloud data of the input inspection features by extracting the point cloud data of the inspection features and automatically labeling the robot inspection trajectories. Meanwhile, the accuracy of the generated trajectories was improved by incorporating a self-attention mechanism module with point cloud geometric features of compliant sheet parts. Then, a diversity loss function was proposed in conjunction with the robot kinematics to improve the diversity of the data generated by the generative adversarial network and solve the solution problem under the uniqueness of the mapping from Cartesian space to robot joint space. Finally, the effectiveness of the algorithm in this paper was verified by case comparison analysis. The results show that the robot inspection planning time is reduced by 52.6% and the end trajectory accuracy is improved by 67.4%.