Robust control of multi-joint robotic arm visual servoing based on neural network and adaptive command filtering

Abstract

This paper proposes a multi joint robotic arm visual servo robust control strategy based on neural networks and daptive command filtering, aiming to solve dynamic uncertainty and nonlinear problems. By constructing an image plane feature point error model and deriving the image Jacobian matrix, a mapping relationship between image velocity and joint velocity was established. Comparative experiments were conducted under standard trajectory tracking, lighting variation, partial occlusion, moving-target tracking, and actuator torque-constrained conditions. The proposed controller was compared with traditional IBVS, fuzzy logic control, LSTM-based control, and ablation variants without the RBFNN, adaptive command filtering, or anti-saturation compensation. The reductions in tracking RMSE and the improvement in tracking success rate were evaluated against the traditional IBVS, fuzzy logic control, and LSTM-based control baselines. The reductions in filtering error and computational complexity were evaluated through comparisons with the fixed-command-filter baseline, while the effect of the anti-saturation compensation was assessed using ablation variants under torque-constrained conditions. Detailed percentage improvements are reported together with the corresponding baselines and test conditions in the experimental section.