Cognitive Systems (eBook)

Preface 6
Contents 8
List of Contributors 10
Part I Introduction 13
Cognitive Systems Introduction 14
Introduction 14
Objective of Project 15
The Problem 15
The Way Forward 15
Steps to Success 15
A Motivating Example 18
Organization of the Research/Research Questions 20
Architecture 21
Representations 24
Learning 26
Perception-Action Modelling 28
Continuous Planning and Acting in Dynamic Multiagent Environments 29
Models of Action and Communication for Embodied Cognitive Agents 34
Multi-Modal Recognition and Categorisation 37
Scenario Driven Research 39
Exploration/Mapping of Space 40
Models for Object and Concepts 44
Consortium 50
Organization of the Book 51
References 52
Part II Component Science 60
Architecture and Representations 61
Introduction 61
Architectures and the Science of Cognitive Systems 62
Requirements for Architectures for Cognitive Robots 64
A New Architectural Schema 67
Key Features of CAS 67
Subarchitecture Design 68
System Wide Design 69
CAST: A Toolkit Implementing CAS 71
Four Problems 72
Binding 73
Filtering 82
Filtering Summary 86
Processing Management 86
The Relationship of CAS to Previous Work on Architectures 98
Cognitive Architectures 98
Robotic Architectures 99
Summary of Contributions and Conclusions 101
References 101
The Sensorimotor Approach in CoSy: The Example of Dimensionality Reduction 104
Introduction 104
Artificial Agents and Human Perception 105
Dimensionality Reduction 107
Review of Classical, “Passive” Approaches 107
Standard Issues 109
The Central Issue of the Metric 111
Dimension Reduction in the Context of Sensorimotor Interactions 114
Finding a Mathematical Framework 115
Back to Dimensionality Reduction 121
An Embodied Algorithm for Dimension Reduction 123
Description of the Algorithm 124
Results 129
Related Work 130
Conclusion 134
References 135
Categorical Perception 140
Introduction 140
Towards Hierarchical Scalable Representations 143
Towards Representations for Affordance-Based Categorization 143
Representations and Discovery of Object Classes by Generative Decompositions 144
Representations of Object Dynamics 144
Low-Level Features and Hierarchical Representation Learning 145
Towards Scalable Representations for Visual Categorization 145
Representations for Functional and Affordance-Based Categorization 152
Mid-Level Representation and Detection 157
Towards Adaptive Representations 157
Learning of Generative Decompositions 157
Generative/Discriminative Hybrid Model for Detection 160
Results on Visual Category Detection 161
Discussion 161
High-Level Representations and Dynamic Models 162
Appearance Model for Single-Frame Detection and Pose Estimation 163
Representing the Dynamics of the Human Walking Cycles with Latent Variable Model 164
Robust Detection and Tracking of People in Image Sequences 165
Outlook and Discussion 167
References 168
Semantic Modelling of Space 174
Introduction 174
Related Work 175
Outline 176
Background 176
Overview of the Spatial Model 177
Metric Map 177
Navigation Map 179
Topological Map 179
Conceptual Map 179
Metric Mapping 180
M-Space 180
Single Camera Bearing Only SLAM 183
Using Visual Attention for SLAM 184
Visual Scans 185
Navigation and Topological Maps 186
Building the Navigation Graph 187
Space Segmentation and Topological Graph 188
Adding Object Information 188
Adding Semantic Place Information 189
Conceptual Map 191
Object Detection and Recognition 193
Object Search and Localization 194
Object Distance Estimation 197
Place Classification 199
Multiple Cues and Modalities for Place Classification 201
Architecture of the Place Classification System 202
Laser-Based Place Classification 203
Vision-Based Place Classification 206
Discriminative Cue Integration 209
Adaptive Place Classification 211
Experiments with Place Classification 212
Single-Cue Place Classification 212
Combining Multiple Cues and Modalities 216
Adaptive Place Classification 218
Semantic Labeling of Space 219
Summary 222
References 224
Planning and Failure Detection 231
Introduction 231
The Multiagent Planning Language MAPL 234
Continual Planning 236
Assertions 236
Assertional Planning 238
Probabilistic Monitoring of Dynamic Processes 241
Sequential State Estimation 242
Particle Filters for Nonparametric Bayesian Filtering 244
Modeling the Influence of Failures Using Hybrid DBNs 245
Gaussian Processes Proposals for Failure Events 246
Data-Driven Proposal Distributions 247
Learning Sampling Models from Data 248
Predicting Collision Events and Parameters 250
Implementation of Sensor-Level Monitoring 251
Evaluation 252
Continual Collaborative Planning 253
MAPSIM 258
Situated Dialogue as Continual Collaborative Planning 260
Related Work 263
Conclusion 266
References 268
Multi-modal Learning 273
Introduction 273
Continuous Learning Framework 275
Introduction 275
Different Modes of Learning 278
Learning Algorithm 279
Reconstructive Representations for Interactive/Online Learning 281
Experimental Results 283
Discussion and Outlook 286
Cross-Modal Learning of Visual Categories 288
Object Representation by Scale-Invariant Patterns 289
Data-Driven Visual Grouping 290
Combining Unsupervised and Supervised Learning 291
Language System 292
Scene Reasoning 293
Label Propagation and Conflict Resolution 294
Learning Complex Actions 295
Introduction 295
Action Representation 296
Inference 299
Learning 301
Experiments 301
Conclusion 303
Functional Object Class Detection Based on Learned Affordance Cues 303
Related Work 304
Affordance Cue Acquisition 305
Functional Object Category Detection 306
Experiments 307
Conclusion and Outlook 308
References 313
Situated Dialogue Processing for Human-Robot Interaction 318
Introduction 318
Background 321
Multi-level Integration in Language Processing 322
Language Processing and Situational Experience 323
Talking 325
Talking about What You Can See 337
Talking about Places You Can Visit 348
Talking about Places 348
Representing Places to Talk about 349
Referring to Elsewhere 350
Understanding References to Elsewhere 354
Talking about Things You Can Do 356
Conclusions 358
References 360
Packing Algorithm 365
Part III Integration and Systems 372
The PlayMate System 373
Introduction 373
System Overview 379
Vision SA 381
Communication SA 382
Manipulation SA 383
Spatial SA 384
Binding 385
Motivation and Flow of Control 388
System Level Control of Information Flow 389
Cross Modal Learning 389
Clarification and Question Answering 393
Mediating between Qualitative and Quantitative Representations 396
Conclusions and Discussion 396
References 398
The Explorer System 400
Introduction 400
Related Work 402
Outline 402
System Overview 403
Navigation SA 403
Object SA 405
Place SA 405
Conceptual Mapping SA 406
The Robot Platforms 406
Spatial Modeling and Reasoning 407
Map Acquisition 407
Acquiring the Conceptual Map 409
Cross-Modal Spatial Knowledge Sharing 411
Planning 416
Scenario: Find Object 418
Conclusions 423
References 425
Lessons Learnt from Scenario-Based Integration 427
Introduction 427
But Is It Implemented? 428
Lessons 429
Integrate Ideas First 430
Integrate People Second 432
Choose Your Tools Wisely 433
Find the Scenario Sweet-Spot 436
Beware the Modularity Mantra 440
Conclusion 441
References 442
Part IV Summary and Outlook 443
Cross-Disciplinary Reflections: Philosophical Robotics 444
Introduction 444
Must an Intelligent Robot Use Language? 446
The Role of the Environment 447
Analysing Requirements Is Very Hard 448
Robotics and Philosophy of Science 449
Ontologies and Laws 449
No “Right” or “Best” Designs 450
A Science of Explosive Diversity 451
Individual Variability 452
The “Designer Stance” in Biology 452
Should Requirements Refer to Laws of Behaviour? 453
Environment-Neutral Requirements and Limitations 454
Forms of Representation 454
Architectures 456
De-fusing Diversity: Towers and Layers 457
Generative Frameworks 457
Subdivision into Towers of Functionality 458
Subdivision into Layers of Functionality 458
The CogAff Architecture Schema – One Small Step 460
Omega Architectures 461
Beyond the CogAff Schema 462
Where Are the Linguistic Mechanisms? 463
Varieties of Compositional Semantics 465
The H-CogAff Special Case 466
Study Trade-Offs Not Special Cases 467
Nature-Nurture Trade-Offs 468
Image-Scene Tradeoffs in Visual Processing 468
Trade-Offs Related to Noise and Uncertainty 469
Requirements for Visual Systems 470
Why Do Perception and Action Need Towers? 470
Multi-strand Process Perception 471
How to Acquire Useful Ontologies 472
Varieties of Complexity Reduction 472
Beyond J.J. Gibson’s Affordances 474
Implications of Speed of Human and Animal Visual Perception 475
Learning to Be a Mathematician 476
Two Kinds of Causation 477
Confusions about the Role of Embodiment 478
Developing the Revolution in Philosophy 479
Further Documentation on These Ideas 480
Why Other Disciplines Need AI 480
Conclusion: The Future 481
References 481
Lessons and Outlook 488
Introduction 488
Lessons 489
Outlook 491
Author Index 493