Data-driven inverse kinematics for a 6-DOF robotic manipulator using an MPGA-optimized BP neural network

Abstract Inverse kinematics (IK) of multi-degree-of-freedom robotic manipulators is challenging due to strong nonlinear coupling, multiple solution branches, and sensitivity to singular configurations. This paper presents a data-driven IK method that integrates a multi-population genetic algorithm (MPGA) with a backpropagation (BP) neural network to improve mapping accuracy and numerical stability. A large training dataset of pose–joint pairs is generated via forward kinematics and Monte Carlo sampling. MPGA performs global optimization of network parameters through Gray-code encoding, adaptive operators, and elitist migration, while the BP model is enhanced with LeakyReLU activation, batch normalization, dropout, and Adam optimization to improve nonlinear fitting capability and regularization performance. Simulation studies on a 6-DOF RM65-B robotic manipulator implemented in MATLAB demonstrate that the optimized model achieves an RMSE of 7.9 × 10 −7 , with mean position and orientation errors of 2.4 × 10 −4 mm and 1.2 × 10 −4 rad, respectively. These errors reflect interpolation accuracy within the sampled workspace rather than general IK performance under uncertainty or model mismatch. Comparative experiments demonstrate superior convergence rates and numerical stability in the simulated environment. These results demonstrate the effectiveness of the proposed MPGA-BP framework for IK solving under offline simulation conditions.