Second Order Sliding Mode Neural Network-Based Visual Servoing of Robot Hands With Fast Manipulation, Including Transition

Strictly speaking, transition tasks such as those executed by robot hands involve free, impact, and constrained motion regimes, with changing dynamics. Impulsive, unilateral constraints arises in the impact regime, which makes very difficult to design a control system. Moreover, algebraic constraints arise in the constrained regime. The trivial approach would be to avoid impact, and to commute consistently ODE- and DAE-based controller, or to impose virtual constraints to model as a DAE system all regimes. In any case, it is required to know exactly the commuting time. In this paper, a very simple control scheme is proposed based on avoiding impact regime, through zero transition velocity from free to constrained motion, therefore impulsive dynamics does not appear. This is possible because we guarantee exactly the time to commute with a novel well-posed finite time convergence scheme, to produce convergence toward any desired trajectory at any given arbitrarily time and for any initial condition. In this way, ODE and DAE dynamics/controllers commute stably. Inertial and gravitational forces are compensated by a recurrent neural network driven by image-based position and force tracking errors, with a decentralized structure for each robot. The network is tuned on line with a second order force-position sliding modes to finally guarantee exponential tracking.