Aiming at the problems of modeling difficulty and weak model generalization ability in complex assemblies, a position prediction method fusing the Improved Dung Beetle Algorithm (GTDBO) and Perceiver model is proposed. First, the ideal assembly is built and the assembly features under six-dimensional perturbation are collected, and a coupled dataset is constructed by interpolation. After that, the Perceiver model learns the nonlinear mapping relationship between feature deviation and positional deviation and optimizes its key hyperparameters with the help of GTDBO. The algorithm combines game-theoretic equilibrium control, adaptive angular perturbation and dynamic foraging strategies. Experiments on the CEC2017 test set show that the algorithm outperforms the comparison algorithms in terms of convergence speed and solution quality. Finally, the model is compared with Jacobian′s method, BP networks, SVR and the original Perceiver model in predictive control experiments on two types of artifacts. The results show that GTDBO-Perceiver achieves MAEs of 6.72×10-2 and 9.96×10-2 on the test set, respectively. Its ability to converge the deviation to the tolerance range within a finite number of control times and to achieve a balanced distribution of errors in defective piecewise scenarios demonstrates a good generalization capability.