Neural Systems for Robotics

Omid Omidvar,Patrick van der Smagt

• Editorial: Academic Press
• Año: 2012
• ISBN:
  9780080925097
• Categoría: Informatica
• Categoría: Reti neurali

Neural Systems for Robotics represents the most up-to-date developments in the rapidly growing aplication area of neural networks, which is one of the hottest application areas for neural networks technology. The book not only contains a comprehensive study of neurocontrollers in complex Robotics systems, written by highly respected researchers in the field but outlines a novel approach to solving Robotics problems. The importance of neural networks in all aspects of Robot arm manipulators, neurocontrol, and Robotic systems is also given thorough and in-depth coverage. All researchers and students dealing with Robotics will find Neural Systems for Robotics of immense interest and assistance.

• Focuses on the use of neural networks in robotics-one of the hottest application areas for neural networks technology
• Represents the most up-to-date developments in this rapidly growing application area of neural networks
• Contains a new and novel approach to solving Robotics problems

1

Neural Network Sonar as a Perceptual Modality for Robotics

Itiel E. Dror, Mark Zagaeski, Damien Rios and Cynthia F. Moss

ABSTRACT

Sonar (SOund NAvigation Ranging) employs a transmitter to generate an acoustic signal and a receiver to register echoes returning from objects in the path of the sound beam. Sonar has been extensively used in robotics for object detection, ranging, and avoidance. However, such applications represent a limited use of sonar, as they do not exploit the full range of information carried by the sonar echoes. This is evidenced by the remarkable perceptual capabilities of echolocating bats, who demonstrate that sonar can convey detailed information about the environment. In this chapter we present data showing that a relatively simple neural network sonar system can perform complex pattern recognition tasks, and propose that sonar has great potential usefulness for robotics. We suggest that sonar has applications in robotics not only for detection of objects and ranging, but also for gathering detailed information about these objects. We begin the chapter with a brief description of the current use of sonar in robotics, and a short tutorial on the bat biosonar system. We then present neural network sonar systems that recognize faces and objects independent of perceptual variations, and a neural network sonar system that can recognize the speed of a moving target based on a single echo.

1.1 Use of Sonar in Robotics

Mobile robots need to navigate in an unpredictable and hazardous world. They must be able to detect obstacles in their path and avoid them. Sonar has been a useful tool for obstacle avoidance; using an ultrasonic ranging system (originally developed by Polaroid for focusing cameras), robots can easily detect the presence of obstacles. The system transmits an ultrasonic pulse from an electrostatic transducer and measures the time it takes for an echo to return. If no echo returns within a given time window, then the robot may continue along its path unobstructed. If, however, there is an echo, then the robot registers the presence of an obstacle and can alter its movement accordingly.

Many robot systems employ sonar for obstacle avoidance. For example, CARMEL (Computer Aided Robotics for Maintenance, Emergency, and Life support) uses a ring of 24 Polaroid sonar sensors to avoid obstacles. However, in this system the use of sonar is limited to obstacle avoidance, and a video camera is used for object recognition.

Polaroid manufactures ultrasonic electrostatic transducers, which can be driven by a variety of modules (e.g., Polaroid sonar ranging module 6500 series, Texas Instruments sonar ranging module SN 28827). A single electrostatic transducer functions as a microphone as well as a speaker; the circuitry generates a broadband chirp which is transmitted through the speaker, and the residual charge left on the capacitance of the speaker provides the high-voltage polarization necessary for it to perform as a condenser microphone and receive the returning echo. The transducer emits a 1-ms ultrasonic chirp, which may contain frequencies as low as 49.4 kHz and as high as 60 kHz. These range-finding modules can work in a single-echo mode, in which only the first returning echo is registered and processed, or in a multiple-echo mode, in which multiple returning echoes (from multiple targets at different distances) are registered and processed.

The Polaroid sonar ranging system uses 12 gain steps on the receiver that are incremented as the echo delay increases; this mechanism allows for a more constant signal-to-noise ratio as the target distance increases and the echo amplitude decreases. The Polaroid system can detect the range of an object from approximately 15 cm to approximately 11 m with an accuracy of 1%. It is low in cost, highly reliable, and easily interfaced with robotics systems. Hence, it provides a convenient tool to give a robot range information about its surroundings.

1.2 Echolocating Bats

Bats are flying mammals that use a biological sonar system to navigate and hunt insects in the dark [9]. As in artificial sonar systems, their biosonar comprises two components: a transmitter (the bat’s vocal apparatus) and a receiver (the bat’s auditory system). The time delay between the emitted sound and its returning echo indicates the range to a target (a use for sonar that is widely implemented in robotics, as described earlier). However, research on echolocating bats has shown that sonar echoes can convey much more information about a target than just its range. Bats can discriminate between targets that differ in shape, size, and movement [6, 29, 31, 32]. There are about 700 species of echolocating bats, and their sonar signals vary widely. The signals used by different bat species appear to reflect the ecological demands on their sonar systems [21]. Some echolocating bat species emit constant-frequency (CF) echolocation sounds, whereas other species use primarily frequency-modulated (FM) sounds.

The big brown bat, Eptesicus fuscus, changes the characteristics of its FM sounds as it pursues an insect, presumably to enhance specific information it is seeking at different points during this pursuit sequence [16, 29]. While it is searching for a target it emits a cry with shallow frequency modulation, sweeping from about 28 kHz down to about 22 kHz in 5–20 ms. When the bat detects a target, it broadens the bandwidth and shortens the duration of its cry. Sounds emitted during this approach phase of insect pursuit consist of a fundamental which sweeps from about 50 kHz down to 25 kHz in 1–3 ms and higher harmonics, yielding a total bandwidth of over 75 kHz. During the final phase of insect pursuit, as the bat closes in and intercepts its prey, the bandwidth and duration of the sonar emission decreases.

By changing the features of its sonar sounds systematically as it approaches an insect, the bat is able to gather the information it needs, while also meeting the changing constraints on its echolocation system (e.g., the shorter delay requires a shorter duration sound to avoid emission-echo overlap [2]). During the search phase, the relatively narrowband, long-duration sounds maximize the energy within a given frequency band, which enhances the likelihood of detecting a faint target echo. As the bat approaches a target and needs to identify it, the broader band sounds can carry more detailed information about the shape of the target in the returning echoes. Finally, when the bat is about to intercept the insect, the short, rapidly repeated sounds provide the bat with brief, frequent samples of the insect’s position, and allow it to compensate for any last-minute evasive maneuvers the insect might make [24]. In addition to these changes in the bandwidth and duration of their echolocation cries, bats also modify the amplitude of their cries as they approach a target, in a manner similar to the Polaroid ranging system described in the last section. Hartley [13] observed that bats decrease the amplitude of their emitted cries at short distances to a target, compensating for increase of the echo amplitude as the distance decreased. For a detailed discussion of the composition of sonar echoes used by echolocating bats for target recognition, see Simmons et al. [30].

1.3 Neural Network Models of Biosonar

Neural networks have been applied to the study of biological sonar. One sonar neural network system was required to recognize a cube and a tetrahedron independent of orientation [4]. The neural network was first trained on a set of 160 echoes of both shapes in different orientations, and then was required to generalize and recognize the shapes from novel orientations. The network reached performance levels of 95% accuracy in recognizing the shapes at novel orientation. The network used broadband frequency-modulated (FM) sounds similar to those used by the big brown bat, Eptesicus fuscus, during the approach phase of target pursuit [10, 16]. In their study, Dror et al. manipulated the input representation to the neural network in order to explore how target shape may be encoded in the echoes. They found that a spectrogram representation of the echoes conveyed the shape information most efficiently. The network was able to recognize the shapes using a power spectrum representation as well; however, its performance was not as high as that when the spectrogram representation was used. Furthermore, by passing the echoes through lowpass and highpass filters, they showed that the first harmonic (with a bandwidth of roughly 50 to 25 kHz) or the second harmonic (with a bandwidth of roughly 100 to 50 kHz) of the spectrogram and the power spectrum representations alone conveyed sufficient information for shape recognition. When time waveform and cross-correlation representations were used, however, the network failed to learn the task altogether.

In another air sonar neural network, more complex shapes were used, as well as a larger number of targets [3]. In this study, neural networks were trained to recognize faces which varied in their facial expressions, and to recognize the gender of faces. After training, the performance of the networks was evaluated on their ability to recognize the faces when they were presented with a novel facial expression (one that was not included during training). The networks were able to learn to recognize five faces with almost perfect accuracy (above 96%) and to correctly recognize gender (with accuracy levels of 88%).

Other neural network studies modeled the underwater sonar system of dolphins, animals that use brief clicks for echolocation. Although underwater sonar operates under different constraints than sonar in air [1], it is worth considering such neural networks as they may provide ideas useful for constructing air sonar neural network systems suitable for robotics. Gorman and Sejnowski were the first to construct such a network [8]. They trained a network to discriminate between underwater echoes from a metal cylinder and a rock and found that increasing the number of hidden units in their network improved the rate of learning the training set. However, such an increase improved the performance on the novel orientations only up to some peak level. After reaching this performance level, further increases in the number of hidden units produced no further impr...

Índice

• Cover image
• Title page
• Table of Contents
• Copyright
• Contributors
• Preface
• Chapter 1: Neural Network Sonar as a Perceptual Modality for Robotics
• Chapter 2: Dynamic Balance of a Biped Walking Robot: Adaptive Gait Modulation Using CMAC Neural Networks
• Chapter 3: Visual Feedback in Motion
• Chapter 4: Inverse Kinematics of Dextrous Manipulators
• Chapter 5: Stable Manipulator Trajectory Control Using Neural Networks
• Chapter 6: The Neural Dynamics Approach to Sensory-Motor Control: Overview and Recent Applications in Mobile Robot Control and Speech Production
• Chapter 7: Operant Conditioning in Robots
• Chapter 8: A Dynamic Net for Robot Control
• Chapter 9: Neural Vehicles
• Chapter 10: Self-Organization and Autonomous Robots
• Subject Index