Teruhiko Ohtomo

Obstacle avoidance travel control of robot vehicle using neural network

This paper describes an intelligent travel control algorithm for a mobile robot vehicle using neural networks, and proposes a method that realizes path planning and generation of motion commands simultaneously. Smooth moving trajectories are controlled by the outputs of cascaded identification modules that have learned the dynamic characteristics of a mobile robot vehicle with strong nonlinearities of both driving force and steering angle. A system is adopted that mutually transforms the absolute coordinate and dynamic coordinate. Because a consequence of the coordinate transformation in this system is that the dynamic position values are normally zero, it is possible to reduce greatly the number of training patterns and, at the same time, to be able to construct an environment similar to that in which a human being drives a vehicle. A travel control system, by which a mobile robot vehicle can move on a smooth traveling path and avoid obstacles, is created by introducing a danger function as an expression of static and dynamic obstacles in an unstructured environment. Finally, the validity of the proposed travel control system is confirmed by computer simulations.

Transfer movement control of mobile robot using two circular arcs

This report describes a method of controlling a robot vehicle to move smoothly on a visually natural path. In this method, discrete points which the robot passes, and directions toward which it moves, are provided on an arbitrary coordinate system given in a movement environment. The continuous curve required for the robot vehicle to move is represented by a total curvature function obtained by accumulating curvatures of curves. Using the function, an algorithm is obtained for generating the robot moving path from interpolation of two circular arcs containing both the lengths and curvatures of the two circular arcs. The position at which the robot is passing and the direction toward which it is moving are measured based on the principle of the triangle survey for three reference patterns set in the movement environment. This is possible by measuring the directions from the current position to the patterns with an ITV camera. This method produces a directional error caused by the roughness of the step angle (0.9) of the stepping motor driving the ITV camera. This paper proposes a method of compensating the directional error by taking into account the number of pixels in pictures taken per step. This system has the averaged position measurement error of 0.8 percent against the coordinate system including the reference patterns. From the result, it is proved that a comparatively simple method enables the moving path of a robot to be controlled accurately.

Domain wall motion in permalloy thin film cores with toroidal and multi-aperture configuration

Domain wall motion and flux patterns in vacuum-deposited thin film cores were observed under dc and pulsed magnetic fields by using the Kerr magnetooptic effect. Two of the thin film cores have a uniaxial magnetic anisotropy around the circumference, and the other is magnetically isotropic. The domain configuration with current through the center of the hole is circular, and expands radially. Domain wall position was quantitatively measured as a function of dc and pulse current. Circular multi-domain walls like an onion slice appear by switching the polarity of the current, and are moved by the pulse field. By exciting the thin film cores with a drive winding, a notched domain appears without continuity of magnetic flux, but the thin film coated on the permalloy bulk core has a circular domain with continuity of flux. From these results, it is confirmed that the thin film has the domain configuration which depends on the distribution of applied magnetic field, and that the thin film coat transcribes the domain wall motion in the bulk core.

Variable‐target‐time travel control algorithm of a robot vehicle by a neural network considering the centrifugal force

The derivation of mathematical models for robot vehicles is extremely difficult due to the high non-linearity of the motion characteristics, making it hard to apply the conventional control theory without using approximation methods which assume constant speed, etc. In this paper, the motion control method using a neural model proposed by Uno, Kawato et al. is extended in such a way to allow its application to the travel control problem of a robot vehicle, which is highly nonlinear with respect to the driving force and steering angle, and a method that generates motion commands and performs trajectory planning at the same time is proposed. In this case, the internal coordinates of the identification module which learns the motion characteristics of the robot vehicle are viewed as motion coordinates, and a method to convert the absolute coordinates for each discrete time into motion coordinates is proposed, making it possible to reduce greatly the number of learning patterns. Furthermore, a variable target time algorithm is proposed that counteracts the centrifugal force effects which act on the robot vehicle when the curvature of the generated trajectory is large. The proposed algorithm restrains the centrifugal force and generates optimal motion commands and trajectory, producing results close to those obtained by a human driver. The effectiveness of the proposed method is demonstrated by computer simulation.

Recognition of three- vowel sequences on the basis of perceptual phoneme boundary lines defined by a flexible model

The authors are attempting to clarify the emphasizing mechanism of phoneme perception in co-articulation. Psychoacoustic experiments were made on the symmetrical and asymmetrical three-vowel sequences, and the effect of the preceding or succeeding vowels on the central vowel is examined. Based on the examined effect, a model is proposed wherein the perceptual phoneme boundary lines of the central vowel are set, considering the three-vowel sequence as a unit. Using this model, the perceptual phoneme boundary line of the central vowel of the three-vowel sequence can be defined by adjusting the 50 percent phoneme boundary line in the perceptual phoneme distribution of the steady vowel on F1-F2 logarithmic plane, according to the positions of the formant frequencies of the preceding and succeeding vowels. A model is presented here which defines the phoneme boundary lines of the central vowel in the three-vowel sequence with co-articulation, by adjusting the 50 percent perceptual phoneme boundary lines of the steady vowel. It is then applied to the real spoken vowels, and its effectiveness is evaluated. Then an automatic recognition system for the three-vowel sequence is proposed, which is based on the trajectory of the three-vowel formant frequency and the perceptual phoneme boundary lines set by the flexible model. A recognition experiment is conducted for the three-vowel sequences using the system, and the usefulness of the system is demonstrated.

Auditory compensation effects in symmetrical three-vowel sequences and a model of perceptual phoneme boundary

To recognize the continuous speech by phonemes, the coarticulation must be analyzed. This paper presents a result of psy-choacoustic experiment for symmetrical three-vowel sequences, which are generated by the terminal-analog-type speech synthesizer, indicating the auditory compensation effects in coarticulation on the phoneme perception. Using each of five Japanese vowels as a reference vowel V0, the symmetrical three-vowel sequences, /V0-VV0/ are formed. The distribution of the perceptual phoneme for the middle vowel is examined first. By comparing the results with the perceptual phoneme distribution for five stationary vowels, the shift of the perceptual phoneme distribution due to the auditory compensation effect is investigated. A model is proposed for the boundary shift in the perceptual phoneme distribution. In the model, the 50-percent phoneme boundary of the perceptual phoneme distribution for five stationary vowels is approximated by straight lines on the F1 - F2 logarithmic plane. The straight lines are contracted in the direction of (F1, F2) point of the reference vowel along the F1 and F2 axis. The proposed model is applied to the perceptual phoneme distributions of the middle vowels in the symmetrical three-vowel sequences, and its usefulness is demonstrated.

Learning and discrimination of perceptual vowel distribution by a neural net model

The phoneme perception distribution is determined by the psychoacoustic experiment for the synthesized speech. The purpose of this paper is to attempt a learning of this distribution by a neural net model with three-layered structure. The following observations are made as the result of learning by a neural net model. The weight of the connection from the input layer to the hidden layer has the function to extract the particular frequency range, and the weight from the hidden layer to the output layer has the function to integrate the unit output in the hidden layer, to extract the particular phoneme. It is shown by integrating those results of analysis that the weight sequence from the input layer to the output layer has the function as a filter to extract only the particular phoneme. Finally, the discrimination power is examined when a part of the connections among units is destroyed or the connection strength of some units is varied. It is shown that the discriminating power is not affected by the change of the weight of the connection between the input layer and the hidden layer, with the effect on the output unit being less than several percent. This result indicates that the neural net model has a feature of distributed memory.