A non-parametric positioning error compensation algorithm for multi-joint industrial robots based on a hybrid whale-genetic algorithm and dual-network cascaded closed-loop dynamic filtering and compensation

Abstract

Multi-joint industrial robots are increasingly used in machining for their flexibility; however, pose-dependent errors caused by nonlinear coupling in low-stiffness structures remain a primary barrier to accuracy. A non-parametric compensation framework is introduced that integrates: (1) a Whale–Genetic hybrid (WGA) global–local optimizer for rapid and stable hyperparameter search; (2) a Median Absolute Deviation (MAD)-based dynamic filter that estimates distributions online and generates adaptive thresholds to suppress impulse and multimodal noise; and (3) a dual-network cascade for “static modeling–dynamic correction,” in which a pose-dependent map is refined by constrained, small-step residual updates in closed loop. On a unified dataset and evaluation protocol, the WGA-BP predictor reduced MAE/MSE/RMSE by 88.4%/98.3%/87.1% relative to a baseline BP and increased R² to 0.995. With MAD and the cascaded compensation, triaxial MAE along a dynamic trajectory decreased from 1.676/2.791/1.783 to 0.094/0.071/0.125 mm. Relative to three baselines—bias-only, axial affine-linear, and third-order polynomial—the iterative closed-loop scheme achieved 85.3%–93.0% reductions in MAE and 79.5%–92.1% reductions in RMSE, exhibiting near half-order convergence, strong noise robustness, and stable generalization. The method provides a practical path from rigid execution to flexible adaptation and supports scalable, cross-platform deployment in advanced manufacturing.