Discretized Data-Driven Neural Dynamics for Model-Adaptive Kinematic Control of Redundant Manipulators

This paper proposes discrete-time learning algorithms that utilize a data-driven technology to address the uncertain issues of optimization and structure. The main challenge lies in acquiring accurate optimization indices and Jacobian matrix, which can be addressed through iterative estimations enabled by these algorithms. On this basis, we propose a new model-adaptive kinematic control (MAKC) scheme for redundant manipulators without prior structure knowledge, incorporating the estimated optimization index and Jacobian matrix. To solve this scheme, a discretized data-driven neural dynamics ( $\text {D}^{3}$ ND) controller is proposed based on the 94LVI algorithm, Kalman filter, and discrete-time learning algorithms. Theoretical analysis is provided to demonstrate its convergence. Subsequently, simulations and experiments are carried out on redundant manipulators using manipulability and joint drift as performance criteria. The results substantiate the robustness, practicability, and superiority of the proposed controller when encountering uncertain issues. Note to Practitioners—Traditional optimization schemes for kinematic control of redundant manipulators may encounter structural and optimization uncertainties due to modifications or environmental factors, which reduce their practicality. To address uncertain issues, this paper proposes discrete-time learning algorithms to estimate the Jacobian matrix and optimization index in real time. Based on the estimated information, an MAKC scheme is designed to achieve kinematic optimization for redundant manipulators without prior structure knowledge. Furthermore, a $\text {D}^{3}$ ND controller is developed, integrating the 94LVI algorithm, Kalman filter, and discrete-time learning algorithms for optimization control, noise resistance, and parameter estimation, respectively. The global convergence of the proposed controller is demonstrated through three theorems. The advantages of the $\text {D}^{3}$ ND controller in addressing uncertain issues are evident from simulative and experimental results when compared to existing methods.