2. Chapter 1: Introduction
“Texas Eastman Co. has collaborated with MCC (Microelectronics and Computer Technology Corporation) to develop an application where they are optimizing the amount of an expensive chemical used as a raw material. The NN (Neural Network) was trained with the historical data from operational records to develop a model of the plant. They claim to be saving almost a third on the raw material and also obtaining higher-quality final product. Texas Eastman plans to explore NN technology for use in all of their plants” [Chemical Engineering, 1990]
The ability of a neural network to estimate an optimum could be applied to robots. Generally, industrial robots do not work at optimal speeds, to achieve minimum motion time. This is because current limitations in computation restrict the current techniques for optimal motion planning. Two of the contemporary ways of dealing with this problem are to i) generate motions that are locally optimal with optimal control techniques, or ii) to plan motions on a global scale, with optimization techniques.
The non-linear characteristics of a manipulator make it hard to model for classic optimal control techniques. Current adaptive control techniques show great promise, but require too much computation time to be successfully implemented. Kuntze , suggests that hardware requirements for sampling speed, processing speed and memory volume are very demanding on model based algorithms. He also says that current complex robot control models require sophisticated computers, experimentation and modelling procedures. Even when these methods are implemented, they still do not necessarily produce optimal motions.
A wealth of optimization techniques makes optimal path planning possible. But, the optimization techniques demand extensive computing resources, and thus may only be performed off-line. When any trajectory is planned off-line, the system must work in an open-loop manner. The open-loop control means that the path performance may degrade in the presence of disturbances and noise.
To summarise, it is very difficult to produce optimal motions in real time, on-line. And, if a motion is planned off-line then the on-line motion tracker may not recover from a disturbance in an optimal manner.
Neuman and Tourassis  have identified some significant engineering tasks as i) Development of methods to solve non-linear robotics equations and ii) Design of computationally efficient algorithms for on-line control. This can be clarified to “the ideal motion planner would be one that incorporates the non-linear dynamics of the manipulator in a real-time optimal path planning scheme.”
Neural networks have been shown to estimate such non-linear problems, and they have the advantage of working very fast, because of an inherently parallel structure. These networks also have other advantages that will be discussed later.
A neural network is capable of taking a set of examples of robot motions, and generalizing to other motions. Thus, the issues at hand are, how to set up the neural network, and what sort of path data is needed for training. Both of these topics will be dealt with in this thesis for different test cases.
Kinematic controllers, based upon maximum velocity and maximum acceleration, will be the first two cases examined. These two cases are examined to show the path planning abilities of the neural networks. After kinematic path planners are shown to be feasible, the problem of dynamics will be explored with a maximum torque controller.
This order of cases is chosen to take advantage of problem features. The maximum velocity motion planner only uses a position error feedback from each joint. The maximum acceleration planner is more sophisticated, and uses a position error and velocity feedback from each joint. This planner has an extra dimension, but it still does not consider the coupling effect of the dynamics between joints. The maximum torque planner uses a full state feedback from joints, and must consider coupling effects between joints.
The results obtained show that neural networks are capable of estimating motions in real time. Although some of these motions are sub-optimal, they exhibit many features of the ‘ideal’ optimal motions. Thus, the results will show that the neural networks are capable of estimating optimal solutions. Finally, recommendations are made for improving the estimates of the optimal solutions.