Learning control and its applications. II: Systems in the Frobenius form and learning control of robots (Q1287282)

[For part I see Autom. Remote Control 56, No. 11, Pt. 2, 1637-1644 (1995); translation from Avtom. Telemekh. 1995, No. 11, 160-169 (1995; Zbl 0923.93059).] This paper proposes a learning control algorithm for control systems in the form \[ y^{(n)}+ f(y,\dots, y^{(n- 1)})= B(y) u, \] where \(u,y\in \mathbb{R}^n\). It is assumed that the matrix \(B(y)\) is invertible and known, while the function \(f\) is Lipschitzian and unknown. The control task amounts to tracking a prescribed trajectory \(y_d(t)\). The main result of the paper consists in the following learning control algorithm \[ u^{k+ 1}= u^k- B^{-1} \Biggl(\mu^{-n}a_{n- 1}\Biggl( \mu{d\over dt}\Biggr) z^{k+1}+ {d^n\over dt^n} z^k\Biggr), \] in consecutive steps \(k= 0,1,\dots\) of length \(T\), where \(z= y- y_d\) denotes the tracking error, \(a_{n-1}(s)\) is an (\(n-1\)) order polynomial such that \(s^n+ a_{n-1}(s)\) is Hurwitz, and \(\mu\) is a small parameter. Conditions on \(\mu\) have been found guaranteeing a convergence of the lerning algorithm in the sense that the error \[ \nu_k= \Biggl(\sum^{n- 1}_{i= 0} \Biggl(\sup_{t\in T}\Biggl(\mu^i {d^i\over dt^i} z^k(t)\Biggr)\Biggr)^2\Biggr)^{1/2} \] tends to zero when \(k\to \infty\). Natural applications of this algorithm to rigid and flexible robotic manipulators have been suggested.