Neural networks and direct vision to position control of fixed eye robotic systems

A vision based position control of robotic manipulator using neural networks to map visual inputs into joint coordinates is presented. Even though neural networks are well suited to learn complex mapping, like this stereo-vision to articular position correspondence, they are approximations and have errors. To avoid imprecisions a feedback control law is proposed in the output space of the neural network (approximate articular positions). Using formal control theory (Lyapunov functions) it is proved a simple well designed feedback control can be stable and convergent, in spite of approximation.<
>