Visual Servoing for Robotic Manipulators: Traditional, Machine-Learning and Deep-Learning Methods, A Literature Review

Abstract

Visual servoing is the closed-loop control of robot motion directly from camera measurements. This review examines how the problem has been solved for robotic manipulators along three successive methodological lines: traditional analytic control, in which the mapping from image features to robot velocity is derived from geometry; classical machine learning, in which that mapping is estimated from motion data rather than modelled; and deep learning, in which perception, and eventually the control law itself, is replaced by trained neural networks. Fifteen widely cited works are analysed in depth, namely the interaction-matrix formalism and the image-based, position-based and hybrid schemes built upon it; online Jacobian estimation and neural reinforcement learning; and deep relative-pose regression, simulation-to-reality visuomotor policies, diffusion policies and vision-language-action models. The three lines are then compared on a common set of axes: accuracy, convergence domain, model and data requirements, computational cost and verifiability. The review concludes that the eras are complementary rather than competing: analytic control still provides the only certifiable inner loop and the best terminal precision, learning provides the perception and semantic breadth that analytic methods never had, and the principal open problem is the disciplined combination of the two.