In this article, the author describes neural network controllers for robot manipulators in a variety of applications, including position control, force control, parallel-link mechanisms, and digital neural network control. These "model-free" controllers offer a powerful and robust alternative to adaptive control.
In recent years there has been increasing interest in universal model-free controllers. Mimicing the functions of human processes, these controllers learn about the systems they are controlling on-line, and thus automatically improve their performance. So far, neural networks have made their mark in the areas of classification and pattern recognition; with this success they've become an important tool in the repertoire of the signal processor and computer scientist. However, the same cannot be said for neural networks in system theory applications.
There has been a good deal of research on the use of neural networks for control, although most of the articles have been ad hoc discussions lacking theoretical proofs and repeatable design algorithms. As a result, neither the control systems community nor US industry have fully accepted neural networks for closed-loop control applications. In this article, I address the major problems facing neural network control and demonstrate that neural networks do indeed fulfill the promise of providing model-free learning controllers for a class of nonlinear systems.
The basic challenges for neural network control are
# providing repeatable design algorithms,
# providing on-line learning algorithms that do not require preliminary off-line tuning,
# initializing the neural network weights for guaranteed stability,
# demonstrating closed-loop trajectory following,
# computing various weight tuning gradients, and
# demonstrating that the neural network weights remain bounded despite unmodelled dynamics-because bounded weights guarantee bounded control signals.
K.S. Narendra and others have paved the way for neural network control by studying the dynamical behavior of neural networks in closed-loop applications, including computation of the gradients needed for backpropagation tuning. (There are also several groups currently analyzing neural network controllers using a variety of techniques.) Unfortunately, the necessary gradients often depend on the unknown system or satisfy their own differential equations. Thus, though rigorously applying them to identification, researchers have not fully developed neural networks for direct closed-loop control.