This paper describes experimental results from a study of neurocontrol applied to a Yaskawa industrial robot. We investigate the use of a hybrid control structure that has superior features and improves the approximation and generalization capability of the neurocontroller. The control structure has guaranteed stability. The results show that the technique has rapid learning speed and excellent performance in a wide range of manipulator tasks. The experiments also demonstrate that the control structure is easy to implement in industrial environments.

We present the first experimental results from a new hybrid learning architecture for maze solving in mobile robotics which attempts to draw on the best ideas from the fields of both “traditional” AI world modelling and behaviour-based robotics. It can operate in both situated geocentric, and nonsituated egocentric modes. In situated mode it learns a “fuzzy cognitive map” of its environment in order to discover a near-optimal path between start and goal position of a particular maze. It is capable of abstracting nonsituated behaviours from a number of such situated learning experiences provided that they share some common features. Then in nonsituated mode it uses the acquired behaviours to navigate through new mazes using only local information

We present a hybrid architecture, called EXP1, which balances exploration and exploitation in order to efficiently solve two dimensional mazes with large state spaces (eg. 262144 states). To achieve this it draws on the strengths of the Genetic Algorithm in search & optimisation, and on the combined strengths of the Radial Basis Function Neural Network and Temporal Difference learning algorithm in approximating continuous functions with strong temporal dependence. The Neural Network acts as an Adaptive Heuristic Critic (AHC). Over successive trials it learns the V-function, a continuous mapping between real numbered positions in the maze and the value of being at those positions. EXP1 solved all the mazes with which we tested it and proved to be quite robust to changes in internal parameters. It also displayed some favourable capabilities in responding to time variant environments.

In this paper we investigate neural network applications in trajectory control of robotic manipulators. Most research in the field remains at an empirical level. Although other authors have claimed very good simulation or even experiment results, lack of theoretical guarantee prevents application of the results in industry. In contrast, this paper presents a neural control method which has a strict theoretical basis. The whole system (manipulator and neural network) stability is guaranteed. Simulations in PUMA robot applications are also presented

This paper discusses the prediction and control of nonlinear discrete systems using neural networks. The discrete systems discussed are neural networks which could be either radial basis functions (RBF) or cerebellar model articulation controller (CMAC). The stability features are guaranteed, i.e. the errors between the predicted values and the actual values in prediction or the errors between the desired values and the actual values in control are bounded. Theoretical results are strict and examples are employed to explain the theoretical results.

The paper extends an online neural network learning algorithm, DBP (derivative backpropagation), proposed by Jin et al (1992) to dynamic systems. The dynamic systems consist of linear systems and backpropagation neural networks. The DBP algorithm learns the desired neural network outputs with respect to neural network inputs. This algorithm increases the position learning speed. Moreover in some neural adaptive control applications the partial derivatives of outputs to inputs are actually used. As argued in Narendra et al (1990, 1991), dynamic neural network systems are very common in control applications, which gives a strong incentive to extending DBP to be a dynamic algorithm

This paper presents a stable neural network control, scheme for manipulators. Cerebellar model articulation (CMAC) or radial basis function (RBF) neural networks are used. The main contribution of this paper is the stability proof of neural network controllers for manipulators. This distinguishes the paper from other work. The results of this paper also have a closer relation to conventional adaptive control. This means that the neural network controller can either work alone if there is no a priori knowledge or work together with conventional adaptive control. Any a priori knowledge can also easily be used to train the neural networks off-line and, therefore, improve the online performance.

The authors present some results on neural controllers. They summarize five neural controller architectures, which can be divided into two classes: general learning and special learning. In the general learning architecture examples, one neural net is trained to simulate the PUMA 560 inverse kinematics and another neural net is used to control the PUMA 560 writing in chalk. The authors then use control theory to analyze special learning architectures. It is found that a simple PID (proportional plus integral plus derivative) controller can provide a momentum factor in backpropagation (BP) learning laws. However, it is found that momentum may make BP convergence unstable, but that one can redesign the PID controller to avoid this problem

An automatic programming method for robot system simulation is introduced. The method integrates `reuse' technology and artificial intelligence to generate the robot simulation program. The overall simulation method for a general system is first proposed. Then a language translator, EXPC, and an expert system, EXPV, are briefly reported. EXPC translates C-based language to C language. EXPV calculates manipulator variables from sensing results. The overall structure of the automatic programming system is presented to show how the system works. Control definition, as an example of a C-based language, is presented