融合物理先验与渐进解耦网络的机器人精度标定OA

To overcome the limitations of traditional geometric calibration in compensating for non-geo-metric errors,as well as the poor interpretability and susceptibility to gradient competition in multi-dimen-sional heterogeneous error fields inherent in purely data-driven black-box models,an absolute accuracy cali-bration method integrating physical priors with a progressively decoupled residual network is proposed.First,a differentiable kinematic grey-box model based on Denavit-Hartenberg(DH)parameters is con-structed as an explicit physical framework to compute the baseline theoretical pose.Second,high-dimen-sional sine-cosine encodings and second-order multiplicative combinatorial features are introduced to en-hance the representation of periodic nonlinear errors.A dual-branch residual network is then employed to independently predict position and orientation residuals,incorporating a differentiable singular value de-composition(SVD)orthogonalization layer to strictly enforce SO(3)manifold constraints.Furthermore,a stage-wise parameter freezing strategy is designed to enable progressive decoupled training,effectively mitigating optimization conflicts arising from the differing dimensionalities of position and orientation.Ex-perimental results on a Staubli TX2-90L demonstrate that the average position error is reduced from 0.377 mm to 0.047 mm.Compared with support vector regression(SVR)and backpropagation(BP)methods,positioning accuracy is improved by 26.3％and 49.9％,respectively.The proposed method achieves a fa-vorable balance between high precision and interpretability,indicating substantial potential for engineering applications such as in situ bioprinting.