Skip to main content

Über dieses Buch

This book is the outcome of the NATO Advanced Research Workshop on Machine Intelligence and Knowledge Engineering for Robotic Applications held at Maratea, Italy in May 1986. Attendance of the workshop was by invitation only. Most of the participants and speakers are recognized leaders in the field, representing industry, government and academic c0mrnunity worldwide. The focus of the workshop was to review the recent advances of machine intelligence and knowledge engineering for robotic appli­ cations. It covers five main areas of interest. They are grouped into five sections: 1. Robot Vision 2. Knowledge Representation and Image Understanding 3. Robot Control and Inference Systems 4. Task Planning and Expert Systems 5. Software/Hardware Systems Also included in this book are a paper from the Poster Session and a brief report of the panel discussion on the Future Direction in Knowledge-Based Robotics. Section I of this book consists of four papers. It begins with a review of the basic concepts of computer vision, with emphasis on techniques specific for robot vision systems. The next paper pre­ sents a comprehensive 3-D vision system for robotic application.



Robot Vision

Robot Vision

This article reviews the basic concepts of computer vision, with emphasis on techniques that have been used, or could be used, in robot vision systems. Sections 2 and 3 discuss two- and three-dimensional vision systems, respectively, while Section 4 briefly discusses some other vision topics. References to basic papers or review papers are given in connection with each topic.
Azriel Rosenfeld

3-D Vision for Robot Applications

Like human vision, robot vision should be able to perceive three-dimensional (3-D) objects, i.e., be able to detect, verify, recognize, locate, inspect, and describe different 3-D objects. Although real objects are three-dimensional, humans can perceive them on the basis of visual data that is two-dimensional (2-D) and incomplete. Utilizing a variety of range cues (e.g., geometric, perspective, texture and shading variations, and occlusion), human perception maps visual data into 3D features and matches them with those of known models. Similar range cues are applicable to robot perception.
D. Nitzan, R. Bolles, J. Kremers, P. Mulgaonkar

On the Computation of Motion from a Sequence of Monocular or Stereo Images — An Overview

Vision and the ability to discern change due to motion are almost universal in the animal kingdom. Moving objects and changing environment surround us. Even stationary objects appear to have (relative) motion because of our own motion or the movement of our eyes. Thus, it is not at all surprising that several distinct types of investigators with widely differing backgrounds and objectives are pursuing research into motion. These backgrounds include psychophysics, neurophysiology, computer vision, computer graphics, and robotics. Briefly, psychophysicists’ and neurophysiologists’ interest in motion centers around understanding the part of biological visual system which senses and interprets motion. Researchers in computer graphics are concerned with the generation of images and, in particular, the generation of moving images and animation on a screen. Researchers in computer vision are interested in the analysis, processing, and understanding of images. A substantial fraction of these researchers are dedicated to the analysis, processing, and understanding of sequences or collections of images with the objective of collecting information from the set as a whole that may not be obtained from any one image by itself. Detection, computation, and understanding of motion are an integral part of these endeavors. Robotics is another discipline where researchers are intimately involved with motion. The current research scene using motion as the common element presents an interesting symbiosis between the disciplines cited above.
J. K. Aggarwal

Time-Varying Image Analysis

We describe two recent projects carried out at the University of Illinois. The first is a computer lip reader, which is used to increase the recognition accuracy of an acoustic word recognizer. The second is a high-level system for representing and identifying time-varying characteristics of a large class of physical events.
Thomas S. Huang

Knowledge Representation & Image Understanding

Knowledge Representation for Robot Vision and Path Planning Using Attributed Graphs and Hypergraphs

This paper presents a general and flexible knowledge representation system using attributed graph representation (AGR) and attributed hypergraph representation (AHR) as the basic data structure. Based on these representations, object recognition and interpretation can be achieved by a hypergraph monomorphism algorithm and a knowledge directed search procedure. A graph synthesis procedure is used to combine the AGR’s or AHR’s obtained from images of different views of an object into a unique AHR. For recognition and location of 3-D objects in 2-D perspective images, another form of AHR, known as Point Feature Hypergraph Representation (PHR) is introduced. With PHR, a constellation matching algorithm can be used to compare images and models as well as to derive 3-D information from stereoscopic images. From the PHR of 3-D objects, procedural knowledge can be formulated and used to search for features in a 2-D perspective image for the recognition and location of 3-D objects in 2-D images. Further, the AGR can also be used to represent the geometric and topological information of the world environment of a mobile robot. A special search algorithm converts the AGR into a AHR from which a compact road map is derived for path and trajectory planning as well as navigation. The proposed method renders greater tolerance to local scene changes.
Andrew K. C. Wong

Knowledge-Based Systems for Robotic Applications

Today, quality design and high productivity in engineering and manufacturing are often synonymous with the use of computers, robots, expert systems, and other computer-based technologies. The greater the degree of computer-based automation exploited and implemented, the greater a nation’s ability to survive in tomorrow’s extremely competitive world market. Among the computer-based technologies, knowledge-based systems for robotic applications is of paramount importance in manufacturing industry.
Julius T. Tou

Image Understanding for Robotic Applications

Current generation robots work in a constrained environment. In most robotic applications, the environment is known and several aspects of it may be controlled. The knowledge about the environment may be used by image understanding algorithms to facilitate the recovery of information from images. Some difficult problems faced by general image understanding systems are simplified by developing techniques that exploit the available knowledge about the environment. We demonstrate the role of such knowledge in robotic applications for recovering information in dynamic scenes. In the first application, the knowledge of ego-motion parameters of a mobile robot is used for segmentation of a scene and recovery of depth information. In the second application, a hypothesize and test approach is used to find road edges in real scenes for an autonomous vehicle.
Ramesh Jain

Robot Control and Inference Systems

Machine-Intelligent Robots: A Hierarchical Control Approach

Intelligent Machines capable of performing autonomously in uncertain environments, have imposed new design requirements for modern engineers. New concepts, drawn from areas like Artificial Intelligence, Operations Research and Control Theory, are required in order to implement anthropomorphic tasks with minimum intervention of an operator. This work deals with the definition of Hierarchically Intelligent Control and the Principle of Decreasing Precision with Increasing Intelligence. A three level structure representing Organization, Coordination and Execution will be developed as a probabilistic model of such a system and the approaches necessary to implement each one of them will be discussed. Finally, Entropy will be proposed as a common measure of all three levels and the problem of Intelligent Control will be cast as the mathematical programming solution that minimizes the total Entropy.
George N. Saridis

On the Application of Intelligent Planning Techniques in Industrial Robotics

This paper explores some of the problems encountered when attempting to apply AI techniques within the domain of industrial robotics. The use of planning techniques is investigated through experiments with two classic AI paradigms. The limitations of these methods areexplored by contrasting the knowledge contained in the case studies with that available in realistic industrial applications. A criticism of the blocks world model is used to highlight the important features of real robot tasks and recommend directions for future development.
M. H. Lee

Analogical Reasoning by Intelligent Robots

Robotics research has produced economically and organizationally satisfactory tools for industry, and exploration of and manipulation in outer space, under the ocean and other dangerous or difficult-to-access places. Intelligent robots, however, are still largely a promising possiblity around the horizon. The adaptation of Artificial Intelligence methodology for robots seems to be a difficult and lengthy process. Both general-purpose and domain-specific techniques are needed. In this paper, we investigate some fairly universal concepts within the block world context.
Analogical reasoning (AR) has long been recognized as an important component of problem solving. In general, AR involves applying the (possibly modified) solution of one problem to a second problem which is in some sense analogous to the first. The prerequisite the two problems have to satisfy is that they have the necessary number and type of important features in common. The task is to discover automatically what the important features are. We discuss at length some general ideas, two basic models and a few advanced processes relating to AR.
Our program generates specific solutions to a number of similar problems that share several properties. The problems are to build certain three-dimensional bodies which satisfy a number of geometrical requirements and constraints. Problem situations are then generalized in the manner of concept formation. Those problems that have similar solutions are replaced with a single concept -- the type definition of a class of problems. Our program, itself, identifies new (hidden or “chunked”) properties it has determined to be essential.
Frames are used to describe problem situations. Four conceptual levels of frames are distinguished: (i) The situation level contains slots for situation properties, the types of available objects, the goal and the eventual solution. (ii) The object level has slots for specific object properties and for lists of possible components that can make up the object. (iii) An unlimited number of component levels look like the object level and represent the components of components…of the objects. (iv) Finally, the property level can contain properties of situations, objects or components.
The underlying learning is a three-stage process. In the first, shapinv stage, heuristic search techniques are used to find a solution to a particular problem. The resulting plan is an action sequence which is then associated with the problem situation. In the second, AR stage, problems with similar action sequences are grouped under a single situation class. A class definition is established which is sufficient to distinguish its members from all other situations. Rules are generated which connect the situation classes and action sequences to be performed in them. The final, consolidation stage compiles the rules into a decision graph. The variables determining the situation class are re-ordered on the decision graph so that the action plans can be retrieved the most efficiently.
Nicholas V. Findler, Laurie H. Ihrig

Knowledge-Based Real-Time Change Detection, Target Image Tracking and Threat Assessment

This paper describes the overall hardware and software architecture of a knowledge-based change detection, target tracking and threat assessment system. It ultimately assigns threat level and threat scenario labels to the observed scene, based on time dependent target features or changes in the scene, as well as target maneuvers and target number. The aims are: reduced scene interpretation workload, shorter reaction time, reduced false-alarm rates, and simple adaptation to changing fields of view/sensors.
L. F. Pau

Task Planning and Expert Systems

Task and Path Planning for Mobile Robots

Autonomous mobile robots represent today a perfect paradigm for third generation robots and possibly the clearest case study for machine intelligence. They imply multi-level environment perception and modelling, decisional autonomy ranging from general planning to specific task operating, autonomous mobility capacity, sophisticated high level man-machine interface, and efficient execution control systems.
This provides for a very large set of research issues as well as for the start up of a broad domain of real-world applications including projects which are outdoors oriented such as ALV and indoors or around-buildings oriented such as HILARE.
Most of the concepts and results presented in this paper are general and can be extended such as to cover man-made and natural-like envionments. We have chosen to use a factory-like environment in which we hypothetically place HILARE, an autonomous mobile robot, as a case study. We also describe very briefly some of the main features of HILARE Mark II, the new experimental robot we have designed.
In this framework we present a world model which includes:
a set of operators, which correspond to elementary tasks, e.g. primitive actions that will be used in the man-machine communication interface.
a three-level environment model: geometrical, topological and semantic
the functional capabilities of the machine, represented by a set of Specialized Processing Modules (SPM) and Specialized Decision Modules (SDM).
The following section presents decision processes organization and inference mechanisms to achieve task and path planning. We stress the importance to produce at this stage an execution model that allows to implement an efficient execution control.
To conclude, in the last section we consider the role of automatic acquisition and learning processes to build and update environment models, and to instanciate and to refine operator parameters.
Raja Chatila, Georges Giralt

Task Planning and Control Synthesis for Flexible Assembly Systems

The development of planning tools which facilitate computer-aided design of assembly systems is important to making the systems implementation process more efficient and cost-effective. Design of flexible assembly systems requires a representation of the assembly task, a planning strategy to select feasible assembly sequences, and a control synthesis procedure to schedule operations and resources. This paper describes a nonlinear planning approach which proposes feasible assembly sequences based on a relational model of part contacts and attachments. The resulting plans are consolidated into an AND/OR graph representation which provides a basis for efficient scheduling of operations. Scheduling efficiency using this representation is compared to fixed sequence and precedence graph representations for a simple example.
A. C. Sanderson, L. S. Homem-de-Mello

Mars: An Expert Robot Welding System

Arc welding is considered to be one of the most promising applications of intelligent robots. This situation first stems from a low manual productivity due to the severe environmental conditions resulting from the intense heat and fumes generated by the welding process. Second. arc welding is the third largest job category behind assembly and machining in the metal fabrication industry.
Pierre Sicard, Martin D. Levine

Software/Hardware Systems

The Edinburgh Designer System as a Framework for Robotics

I discuss how the capabilities of the Edinburgh Designer System can be extended and used to support symbolic computation for robotics. I conclude that the Algebra Engine requires to handle temporal constructs, groups and tolerances, that the taxonomy can support activity modules and that automatic plan formation would require the creation of a specialist.
R. J. Popplestone

Implementation of Complex Robot Subsystems on Distributed Computing Resources

An approach to the coordination of a complex robot’s subsystems has been developed and implemented. In this approach, the computational load is distributed functionally over several microprocessor systems in both tightly and loosely coupled configurations. Tightly coupled functions communicate through shared memory on the same high speed parallel bus. Loosely coupled functions communicate through a local area network. However, whether tightly or loosely coupled, communications between functional modules appear as if a single blackboard memory is shared. This approach to robot integration has been used to explore various concepts for sensor data fusion. An autonomous mobile robot has provided the experimental environment in which experience with this approach has been gained. The concepts fundamental to this approach have also been extended to coordinate multiple interacting robots.
S. Y. Harmon

Autonomous Robot of the University of Karlsruhe

This paper discusses the autonomous research robot which is being developed at the University of Karlsruhe. The device will perform simple assembly operations in the laboratory. The purpose of the project is to develop new technologies for advanced robotic machines for industrial application. The robot contains a mobile platform, a complex sensor system, two manipulators, hierarchical controls and an expert system. Programming will be done by task-oriented instructions. A considerable amount of the fundamental technology which will be integrated in the robot has already been developed at the institute, including a vision system, a robot hand containing 5 different sensors, a programming system and a hierarchical computer architecture.
Ulrich Rembold, Rüdiger Dillmann

Poster Session

Robotics Research at the Laboratory for Intelligent Systems National Research Council of Canada

This paper summarizes the research and development activities in the field of intelligent robotics at the Laboratory for Intelligent Systems of the National Research Council of Canada. The Council’s objective is to use its long term research to provide a base of knowledge from which to attack short term problems in support of the Canadian industry. The Laboratory’s work on three-dimensional vision, sensory based control, multiprocessor system architecture and applications of artificial intelligence is presented.
S. Elgazzar

Future Directions in Knowledge-Based Robotics

Future Directions in Knowledge-Based Robotics

On the final day, the participants gathered to discuss and summarize the major issues raised and to look at future research directions. Each of the six sections of the panel discussion was aptly led by a researcher who first reviewed the observations made during the workshop.
Deborah A. Stacey, Andrew K. C. Wong


Weitere Informationen