This chapter describes the design and implementation of a localisation scheme for ARNE. Without such a technique ARNE's estimated position, based only on odometry, would gradually diverge from its true position.

The essence of localisation is to match recent sensory information against prior knowledge of the environment. Some researchers build a ‘local’ map from the latest sensor readings and then look for the best match between the local map and a global map. The correspondence can then determine the robot's position in the global co-ordinate system. Elfes (1989) does this by seeking the best correlation between local and global probabilistic grids. Crowley (1989) and Drumheller (1987) both extract line segments from the sensor data and compare the position, orientation and length of the each line with lines in the global model.

The experiments in Chapter 6 showed that it is impossible to determine either the type or position of environmental features from a single scan of ARNE's sensor. It is therefore not possible to construct a local map from each viewpoint. Instead, the technique adopted in this thesis is to find immediate correspondences between sensor readings and known features, and to use the range readings to known objects to estimate ARNE's position. Published examples of this approach include Curran (1993), Leonard and Durrant-Whyte (1992), Rencken (1993) and Kleeman (1989). The process of matching ARNE's sensor readings to known objects has already been described in Section 7.3.

ARNE's application requires it to follow efficient paths to user-specified delivery points. This chapter describes how these paths are planned.

Path planning serves two purposes in this thesis. First, it is obviously necessary to move during exploration and, although some of these movements (e.g. during wall-following) may be completely reactive and not use the map, others will require ARNE to go to a specified viewpoint while avoiding known obstacles. These movements will need to be planned.

Another, less obvious, need for path planning is in the measurement of map quality. As will be seen in Chapter 10, map quality is measured by predicting how successful ARNE would be at a number of test tasks, given the latest map. Path planning is needed to make this evaluation.

Section 7.4 described the construction of a free-space map from the list of confirmed features. Path planning is based totally on this map.

The planning technique used in this thesis was first presented by Jarvis and Byrne and is described by McKerrow (1991). A ‘distance transform’ is calculated which indicates, for any given cell in the free-space map, which of the neighbouring cells is closest to the goal. This information can be used repeatedly to generate a list of cells through which the robot can reach the goal.

Section 8.1 gives a brief overview of the technique and Section 8.2 gives the implementation details.

The paths derived from the distance transform are often unnecessarily jerky, giving a zigzag path to the goal.

Figure A.I is a simplified entity-relationship diagram which shows the main components of the feature-based map and the relationships between them.

Tables A.I to A. 11 list the data elements owned by each of the entities in Figure A.I. Note that all of the entities are contained, either directly or indirectly, within the “Map” entity. This reflects the fact that the map is implemented as a single shared data structure within ‘C’.

This book is the product of my PhD research at University College London. I am grateful to the many people and organisations that have made my research both possible and enjoyable. I have benefited greatly from the companionship and support of all of my colleagues in the Computer Science and Anatomy departments during the course of this work. The following paragraphs can only recognise some of the more direct contributions.

Many thanks are due to Michael Recce for his enthusiastic and constructive supervision. Michael has been generous with his time, his ideas, and his lab space. I am also indebted to Michael for his careful reading of draft versions of this document and his valuable suggestions about its style and content.

Jim Donnett has made many much-appreciated contributions to this work, ranging from hardware design and debugging through to a detailed reading of the thesis. Jim's breadth and depth of knowledge have been invaluable, and his friendship and sense of humour have helped me through some trying moments.

I owe a great deal to Clive Parker for the construction of my robot, ARNE. Thanks to Clive's electronic and mechanical skills, a loose collection of components was transformed into an effective research tool.

David Brown of the Statistics Department kindly took the time to advise me about the statistical analysis of my results, despite having been ‘Volunteered’ for the job. His insight and suggestions were most welcome.

17.1 Motivation

The exploration strategies presented so far in Part III of this thesis have differed in the extent to which the map has been used to control the navigational choices. Wall-Following (Chapter 12) and Longest Lines (Chapter 16) were both totally reactive, not using the map at all. Supervised Wall-Following (Chapter 14) used the map to detect circumstances in which wall following was becoming ineffective. Chapter 15 showed the results that could be obtained when a human operator used the map to guide the exploration. This chapter will present an exploration strategy in which ARNE's decisions are driven primarily by the information present in the partially-formed map.

The implementation described in Section 17.2 builds on the ideas presented in Section 4.2; ARNE approaches the interesting regions of the environment. The central issue is, of course, the definition of ‘interesting’. The definition adopted here focuses on the edges of free space, the regions in which free cells are next to unknown cells.

Section 17.3 presents the results of experiments to evaluate this strategy and Section 17.4 summarises the results.

17.2 Implementation

The first step in this implementation was to identify the cells on the free-space map which were to be examined. The decision was made that ARNE should approach unknown regions but, to avoid collisions or panic stops, it should not actually enter unknown regions. The interesting cells are therefore those on the boundary between free and unknown space.

The experience gained during the development of this thesis has suggested a number of directions in which the research could be extended. This final chapter examines these ideas under four groupings:

1. • Mixing planning and reactive navigation.

2. • Modifying the exploration method as the exploration progresses.

3. • Testing new sensors and new environments.

4. • Examining the feature map for inconsistencies.

A section is devoted to each of these areas.

20.1 Mixing Planning and Reactive Navigation

The Supervised Wall-Following strategy has shown that effective exploration can arise from a combination of reactive and model-based decisions. The application of the quality metric to maps of the ‘Walls’ environment showed that small errors in the map could lead to collisions unless the robot's movements took into account the latest information from the robot's sensors. These results suggest that it would be useful to extend the current research by implementing a navigation strategy which combines planning and reactive components.

There are clear parallels between this idea and the concept of compliance in automated assembly (McKerrow 1991, page 293). In both cases the robot uses its stored understanding of the state of the world to plan its actions, but it has to adjust its behaviour if sensory input disagrees with that understanding.

The work of Pay ton, Rosenblatt, and Keirsey (1991) is attractive in this context. They propose the use of ‘internalized plans’ which act as information resources to guide the reactive behaviour of the robot.

This thesis has described an investigation into the complementary problems of map-building and exploration by a mobile robot. This chapter highlights the most significant results of this investigation.

The novel contribution of this research can be summarised as:

• • The integration of a physical robot, a sonar model, map construction algorithms, and a localisation algorithm into an effective working system;

• • The definition and implementation of a novel quantitative measure of map quality;

• • A thorough quantitative and statistical evaluation of the map-building and exploration capabilities of the system, using the quality metric and a variety of exploration strategies.

The system components and the quality metric were described in Part II of this thesis. Sections 19.1 to 19.4 briefly review these topics. The experimental evaluation of exploration strategies formed the bulk of Part III of this thesis. The results of this work have already been summarised in Chapter 18.

Chapter 2 described the continuing debate between the ‘traditional’ supporters of modelbased robotics and the proponents of behaviour-based robotics. An outcome of the current research has been an awareness of the need to balance these two approaches. The value of reactive navigation became more apparent as the research progressed. Section 19.5 reviews the course of the research in the context of the ‘models versus behaviours’ debate.

19.1 The Ultrasonic Sensor Model

Chapter 6 presented a set of experimental results which showed how the Polaroid ultrasonic rangefmder detected each type of object that would be encountered in the test environments.

Chapter 12 showed the loss of map quality which arises as odometry errors accumulate and ARNE's position estimate becomes increasingly inaccurate. Chapter 9 presented a method by which ARNE can improve its position estimate by measuring the distance to known objects in its environment. The following sections describe the implementation of this localisation method and show the results of experiments to test its effectiveness.

A key component of the localisation method, the plant model, models the growth in positional uncertainty as ARNE moves. The plant model requires parameter values which are specific to the individual robot. Section 13.1 describes experiments to check that the parameters were approximately right for ARNE.

Section 13.2 then repeats the experiments from Chapter 12, but this time with the localization system in place, and compares the results. The loss of quality in the later stages of exploration is eliminated.

After the benefits of localisation have been demonstrated, Section 13.3 presents the results of wall-following with localisation in two other, more complicated, environments. The quality is shown to increase more slowly and to reach a lower maximum value in more cluttered environments. The reasons for this loss of quality are discussed.

The results of wall-following are then used to determine the best value for one of the central parameters of the map construction process, the confirmation threshold. Section 13.4 describes the experimental basis on which this choice is made.

Chapter 2 described numerous maps which have been used by mobile robots. This chapter considers which type of map to use in the current research.

The choice of map type is strongly constrained by the proposed application of the robot. In Chapter 1 a delivery application was chosen. Section 3.1 describes such an application in detail.

Section 3.2 uses the knowledge of the application to choose the maps to be used in this thesis. One of the most important choices was between probabilistic grid-based maps and feature-based maps. Section 3.3 explains why feature-based maps were selected. The chapter concludes in Section 3.4 by explaining why the robot will build its own map, instead of being given one by its operator.

The details of the map construction algorithm can not be described without knowledge of the robot and its sensors. This description is therefore postponed until Chapter 7 to follow the descriptions of the hardware and the sensor model in Chapters 5 and 6.

3.1 The Application

The choice of world model is strongly influenced by the proposed application of the robot. Indeed, as was discussed in Chapter 2, some applications do not require a world model at all. It is therefore vital to be precise about the intended application of one's robot before designing the world model.

This thesis addresses the construction of maps for use in an application with the following features:

Chapters 12 to 17 will examine individual exploration strategies and compare their results. This chapter introduces this part of the thesis by examining some general issues which are important whichever strategy is being tested.

Figure 11.1 shows the possible links between exploration algorithms and the rest of the system software. Two extreme types of exploration are represented on the diagram. The top portion of the ‘Explore’ box depicts reactive exploration in which the movement commands are based solely on the most recent sonar readings and the result of the previous command. In contrast, the lower portion depicts exploration which is totally map-driven. The experiments described in this part of the thesis investigate the potential benefits of striking a balance between these two extremes.

What would it mean to say that one exploration strategy is better than another? A reasonable interpretation would be that the first strategy produced a higher quality map than the second, for the same cost of exploration. One then has to decide how to measure the ‘cost of exploration’. Section 11.1 considers some alternatives and selects ‘the total time taken by the robot's movement and sensing actions’.

To make a fair comparison, the strategies must be tested in a variety of circumstances. The effectiveness of a strategy can depend on the environment being explored and on the starting position of the robot within that environment. Each of ARNE's strategies is therefore tested in at least 3 different environments and from 10 starting positions spread throughout each environment.

This chapter summarises the results of the experiments that were reported in Part III of this thesis.

Section 18.1 examines the performance of the sensor model by analysing the results of the wall-following explorations that were presented in Chapter 13. Section 18.2 uses the same experimental data to consider the effectiveness of the feature-based map-building algorithm.

Chapters 12 to 17 described a variety of exploration strategies and presented the results of experiments to evaluate those strategies. Section 18.3 collects those experimental results together in order to compare all of the autonomous strategies across the set of test environments.

Section 6.3 introduced a sensor model in which adjacent sonar returns of similar range were grouped into ‘readings’ to decrease the uncertainty caused by the width and uneven strength of the sonar beam. Table 18.1 shows the number of readings of each size that were taken during the wall-following explorations of the three main test environments. It can be seen that approximately 20% of the readings contained two or more returns. Expressed in terms of the raw returns instead of the readings, the results show that 37.5% of the returns were included into groups of size two or more. (Maximum-range returns are excluded from the grouping process and from this calculation.) This supports the opportunistic nature of this approach, grouping the returns where possible but using all of the available information. An insistence that groups contain at least two returns would have eliminated almost two-thirds of the returns.

Map construction is an essential component of the research reported in this thesis. This chapter examines the reasons why a mobile robot might need a map and reviews the variety of types of world model which have been devised and implemented by previous researchers.

Early research work into mobile robots (Moravec 1983; Crowley 1985) took it as axiomatic that an effective mobile robot would need an environment model. The process of control was viewed as two steps: first the robot uses its sensors to build a world model and then it uses the world model to plan and execute its actions. The details could vary (different sensor modalities, different data structures for the world model) but the underlying two-step process was not questioned.

In the mid-1980's a number of researchers, most prominent among them being Rodney Brooks (1986), became frustrated with the perceived slow progress in mobile robotics and began to search for an alternative approach to the ‘traditional’ dependence on environment models. The intention was to minimise the processing between sensing and action. Robots were built in which there was an almost immediate link between the robot's sensors and its actuators. (Braitenberg's excellent book ‘Vehicles’ (1984) describes, in the form of thought experiments, what could be achieved by such robots.) The robots were able to perform tasks such as approaching beacons, avoiding obstacles, and following walls. These behaviours were found to be very robust. Brooks's robots could operate in unmodified office environments, sharing their world with unpredictable humans.

Published work shows a variety of approaches to exploration for mobile robots, ranging from disregarding the issue completely through to detailed mathematical analysis of exploration algorithms. This chapter reviews this work in the context of the recent debate between ‘reactive’ and ‘model-based’ robotics (as discussed in Section 2.1).

Many of the published papers on the map-building and navigation of mobile robots do not consider the question of exploration at all. This is, of course, often just a choice of research focus; effort is expended on the mechanics of map construction from sensor data without worrying about how the sensing positions were selected. On the other hand there are theoretical reasons why some researchers have chosen not to study exploration. A robot will not need to explore if its application is such that it does not need a map (Brooks 1990, pages 8–9) or if the map is to be supplied by the operator (Crowley 1985; Drumheller 1987). Neither of these arguments apply in the context of this thesis. Section 2.1 argued that a map was needed for the proposed delivery application and Section 3.4 explained the reasons for allowing ARNE to build its own maps.

Some researchers (Engelson (1992), for example) have adopted a strategy of ‘passive’ mapping, in which the map is built while the robot carries out its normal activities. In contrast, the current research proposes an initial exploration period during which the robot' objective is simply to learn about its environment.

Chapters 12 and 13 showed the results of exploring by wall-following. How, if at all, can that performance be bettered? This chapter begins by describing, in Section 14.1, some circumstances in which wall-following appears to be inefficient and then, in Section 14.2, proposes a new strategy, Supervised Wall-Following, to eliminate these inefficiencies.

Section 14.3 presents the results of experimental tests of Supervised Wall-Following. Its value is shown to be higher in more complex environments.

Section 14.4 summarises the results and suggests some directions in which the algorithm could be developed.

14.1 Shortcomings of Wall-Following

In the wall-following experiments described in the previous two chapters, ARNE was specifically denied access to the map while making navigational decisions. A consequence of this was that a human observer watching the exploration (and looking at the map) would become frustrated by ARNE's inflexibility. In certain circumstances ARNE would make movements which, to the human observer, would simply appear to be ‘stupid’. Three of the most obvious circumstances are: falling into traps, re-examining known objects, and repeating fruitless examinations. The remainder of this section considers each of these problems in turn, illustrating with examples.

Figure 14.1 shows a simple example of a wall-following trap. ARNE began at position 6 in room ‘Columns’, close to one of the free-standing cylinders. It began to circulate around the cylinder and continued to do so for the entire exploration period. This obviously restricted its view of the environment and limited the quality of the map.

This chapter describes the map-building algorithms which were designed and implemented in this research. Figure 7.1 shows that the sonar readings (as described in Chapter 6) and knowledge of ARNE's position are combined to generate a sequence of feature-based representations and a free-space map.

The sensor model that was developed in Chapter 6 showed that positional uncertainty in the sonar returns could be decreased by grouping multiple returns from the same viewpoint. But uncertainty, especially angular uncertainty, still remains. The objective of the algorithms in this chapter is to reduce the uncertainty further and build an accurate representation of the environment.

More value can be derived from a sensor reading by examining it in the context of preexisting information about the world. This information comes from two sources; either the latest map or sensor data which has not yet contributed to the map. As an example of the first type, imagine that the current map shows that there is a wall directly in front of the robot. If a sensor reading is then obtained which is consistent with a sonar reflection from that wall, the range reading can be used to update the estimated position of the wall. The uncertainty due to beam width has been eliminated and the new range reading can be averaged with the existing knowledge to limit the impact of unpredictable errors.

Unfortunately map-building is not quite this simple. First, there is the ‘bootstrap’ problem of gathering enough initial information so that sensor readings can be matched with known objects.

Abstract

This article considers the major change that has taken place in natural language processing research over the last five years. It begins by providing a brief guide to the structure of the field and then presents a caricature of two competing paradigms of 1980s NLP research and indicates the reasons why many of those involved have now seen fit to abandon them in their pure forms. Attention is then directed to the lexicon, a component of NLP systems which started out as Cinderella but which has finally arrived at the ball. This brings us to an account of what has been going on in the field most recently, namely a merging of the two 1980s paradigms in a way that is generating a host of interesting new research questions. The chapter concludes by trying to identify some of the key conceptual, empirical and formal issues that now stand in need of resolution.

Introduction

The academic discipline that studies computer processing of natural languages is known as natural language processing (NLP) or computational linguistics (the terms are interchangeable). NLP is most conveniently seen as a branch of AI, although it is a branch into which many linguists (and a few psycholinguists) have moved. In Europe, NLP is dominated by ex-linguists but this is not the case in the USA where there is a tradition of people moving into the field from a standard computer science background.

It is tempting to say that NLP is the academic discipline that studies computer processing of the written forms of natural languages. But that would be misleading. The discipline that studies computer processing of the spoken forms of natural languages is known as speech processin or just speech.