Neural network based repetitive learning control of robot manipulators

Abstract

Control of robot manipulators performing periodic tasks is considered in this work. The control problem is complicated by presence of uncertainties in the robot manipulator's dynamic model. To address this restriction, a model free repetitive learning controller design is aimed. To reduce the heavy control effort, a neural network based compensation term is fused with the repetitive learning controller. The convergence of the tracking error to the origin is ensured via Lyapunov based techniques. Numerical simulations and experiments are performed to demonstrate the viability of the proposed controller.

A learning control algorithm for periodic robot synchronization: Experimental results

“…Such joint position reference signals, which are "periodic" with a time-varying increasing period T ranging in the compact set [2.5, 3.5] seconds, are used to experimentally test the online period identification capabilities of the recursive algorithm (4), (5) in the learning context of (2), (6).…”

“…Theorem 1. Consider the robot dynamics (1) and the learning control (2), (6) in conjunction with the period identifier (4)- (5). For sufficiently smallT(0) and̃(0): (1) the uncertain period T is exponentially estimated; (2) the error vector [q(t) Tq (t) T ] T is exponentially attracted into a ball (with center as the origin) whose radius can be made arbitrarily small by increasing the smallest among the learning gains k Li , i = 1, … , n; (3) the error vector [q(t) Tq (t) T ] T asymptotically tends to zero.…”

“…Remark 1. When compared to the period identifier in the works of Verrelli 15,16 in which a global, finite-time converging, explorative search-based period identifier is presented § (see the other work of Verrelli 8 for its use in an electric motor application), the period identifier (4)- (5) does not resort to the infinite number of evaluations there theoretically involved, although only local estimation is here guaranteed. Furthermore, a recursive period identification solution is here considered, which may easily face to online period time variations (see the subsequent Section 4).…”

“…In particular, assume that the constant period T of the (sufficiently smooth and periodic) joint position reference signal q ⋆ (t) ∶ (−∞, +∞) → R n is known and that the joint position vector q(t) and its reference signal q ⋆ (t) are available at runtime t along with their time derivatives. As shown in the works of Dixon et al 1 and Bifaretti et al, 2 the corresponding control problem is definitely solved by the state-feedback repetitive learning control algorithm in the work of Dixon et al 1 (see extensions concerning finite memory implementations in the works of Bifaretti et al, 2 Salis et al, 3 and Verrelli et al 4 and the extension concerning the use of neural networks in the work of Cobanoglu et al 5 ).…”

Robust Vibration Control Based on Rigid-Body State Observer for Modular Joints

“…Nevertheless, it is time-consuming and computationally complex. Furthermore, limited data information may lead to bad control performance [16][17][18].…”