Active Sensors for local Planning in mobile Robotics
The goal of realising a machine which mimics the human ability to refine and structure behaviour in a complex, dynamic world continues to drive mobile robot research. Central to such ability is the need to gather and manipulate rich information on the surroundings. Such a grand ambition places stringent requirements on the sensing systems and on the interaction between sensor and task.
One thing which has become clear in attempts to achieve this is the need for diversity in sensing systems. The human vision system remains the inspiration for artificial analogues, but none can approach its sophistication in terms of hardware or processing. Structured light systems, which measure range directly through using a light source to probe a specific area, are a more reliable method for artificial planning. Their equivalent in sound, sonar, has increased in adaptability and reliability, driven by collaboration with bat biologists as well as from the more standard and established radar literature. Radar itself is becoming cheaper.
Given such diversity, another requirement is a structure and methodology to share and optimise information. Two important paradigms have arisen as a result. One is the idea of the logical sensor which hides the details of the physical sensing operation, so sensors may be specified in terms of task and not in terms of technology: hence a task might require, for example, a sensor to find line segments under particular conditions, rather than a particular technology such as sonar. The other is the active sensor, which abstracts and selects information according to demand - whether this is through probing the environment physically - for example through emitting radiation (the traditional active sensor) or through choice or tuning of algorithms. This concept is an extension of the traditional formulation of the active sensor which interacts with the environment through emitting radiation such as sound or light. By developing sensors within this framework we avoid the bottleneck of a large information repository.
Much of the work in this book is the result of research with which the editor has been associated in Oxford. It is designed both to provide an overview of the state of the art in active range and vision sensing and to suggest some new developments for future work. It describes real systems and sensors. Cross references have been included between chapters to develop and relate concepts across and within a single sensing technique.
The book starts with a brief overview of the demands for local planning, discussing the problem of finding a reliable architecture to handle complexity and adaptability. It describes the concept of the active sensor, driven by the task in hand and filtering information for that task, to provide a fast, tight sensing-planning loop. It gives an overview of common sensing technologies.
In mobile robots, a key requirement for planning is to find out where the robot is within a known region - the localisation problem. Mapping, the problem of extracting geometric or feature based information often underlies this. Reliable mapping and localisation requires robust and versatile sensors, and also a systematic method to handle the uncertainty inherent in the sensors and in the robot's own position. Chapter 2 addresses generic issues in mapping and localisation and introduces an important algorithm which is referred to many times in the book, the extended Kalman filter.
Sensors which measure range directly are particularly useful for planning. Sensors active in the traditional sense are most important here and most of the book deals with hardware and algorithms for the two most common classes of these: sonar sensors and optoelectronic sensors.
The essential factor which distinguishes the way sensors in these classes view the world is their wavelength. Whereas the data from optical sensors naturally falls into standard geometric descriptions such as lines, corners and edges, millimetre wave sensors such as sonar see the world rather differently. Part II of the book discusses millimetre wave sensors. Significant interpretation is required to extract data for comparison with a standard geometric model. In spite of this, sonar is the commonest sensor used in robotics, largely because of its low cost and easy availability. Another sensor which operates in the millimetre band is high frequency radar - more expensive but with very long range and so of great interest outdoors. Although one of these sensors emits sound waves and the other electromagnetic waves, because of the similar wavelength their data has many similar characteristics. Chapter 3 discusses generally how these characteristics depends on both the sensor geometry (especially the antenna) and target type.
Sonar has seen particular developments in the last ten years, from a simple sensor used for obstacle avoidance to a sensor which will produce reliable and robust maps. Chapters 4 to 6 describe how this has been achieved through advances in hardware and data interpretation. Methods of modulation and signal processing drawn from underwater sonar and military radar have been applied to improve resolution and hence extend the range of environments in which sonar operates (chapter 4). Surface modelling, especially the incorporation of rough surface models, has led to better mapping and application in texture recognition (chapter 5). Drawing on analogies from biology, bio-sonar has improved efficiency through sensor placement and small sensor arrays (chapter 6). Finally the application of new processing techniques, especially morphological filtering, has led to the possibility of curve fitting, to produce information which is geometrically similar to our own perception of the world (chapter 7).
The problem with sonar is power; the maximum range is limited to around 10m or less (normally closer to 5m). Milimetre wave radar has many similar characteristics but will see over ranges huge by robot standards - over several kilometres depending on weather conditions. For this reason it is of great interest in the field, and the increasing use by the automobile industry (for automatic charging for example) means that the cost is falling, although it is still an expensive technology. Chapter 8 describes the capabilities of radar with a summary of some recent work in robotics.
Part III describes sensing at optical wavelengths. Optoelectronic sensors probe the environment using a laser or focussed light emitting diode. At their best, they provide data of high quality which is easy to interpret in terms of standard geometry. However difficulties arise from strong ambient light levels as the active light source can be swamped. A further difficulty in actually realising these systems in the laboratory is the need to scan over one or two dimensions. Unlike scanned sonar, which is compact and light, a scanning optoelectronic sensor imposes power and weight demands which place restrictions on its speed and reactivity. Because of this most applications in local planning gather only two dimensional data (often range versus orientation). Some of these issues are discussed in chapter 9, which also describes some common optical methods to measure range. Chapter 10 describes in detail a sensor based on a technology which has been of particular importance in robotics, amplitude modulated continuous wave (AMCW) operation, often known as lidar. The following chapter (chapter 11) describes the extraction of lines and curves from this and other types of optical range sensor. Chapter 12 describes active vision, in a system which allows the camera to select features of interest and to maintain these in the centre of its field of view through a multi-degree of freedom head. It is impossible to do justice to such an important subject in a book of this scope and it is hoped that this chapter, besides describing a state of the art system for mapping and localisation, will encourage the reader to pursue more specialised texts.
The final part of ths book, Part IV, considers some general issues in sensor management. Chapter 13 describes a system which is showing real benefits for processing visual and infra red data. In addition it introduces the more abstract areas of adaptive sensor and knowledge representation.
The ultimate goal of autonomy remains elusive, but there are many examples of systems influenced strongly by robotics research. Bumper mounted sonar has been introduced as a parking aid in cars; radar is common not just for speed detection but for automatic charging. Surveillance systems draw on active vision to process and abstract information. The multi-agent paradigms used for routing in Internet access have their counterparts in behavioural robotics. The demand for indoor localisation has expanded into areas such as environmental monitoring as a response to the availability of GPS outdoors.
The developments described in this book are relevant to all those who are looking for new and improved ways to handle task orientated information from sensors. It is directed at a final year undergraduate or first year postgraduate level, as well as being of use as a source of ideas to researchers and interested practitioners. Inevitably it has only been able to cover some of the work going on in the field. However I have enjoyed the opportunity to put this book together and I hope that the reader will capture some of the excitement of our research and will use the bibliography as a springboard for their own further investigations.
Subscribe to:
Post Comments
(
Atom
)
No comments :
Post a Comment