Christensen Cognitive Systems

1. Auflage 2010
ISBN: 978-3-642-11694-0
Verlag: Springer
Format: PDF
Kopierschutz: Adobe DRM (»Systemvoraussetzungen)

E-Book, Englisch, 494 Seiten

Design of cognitive systems for assistance to people poses a major challenge to the fields of robotics and artificial intelligence. The Cognitive Systems for Cognitive Assistance (CoSy) project was organized to address the issues of i) theoretical progress on design of cognitive systems ii) methods for implementation of systems and iii) empirical studies to further understand the use and interaction with such systems. To study, design and deploy cognitive systems there is a need to considers aspects of systems design, embodiment, perception, planning and error recovery, spatial insertion, knowledge acquisition and machine learning, dialog design and human robot interaction and systems integration. The CoSy project addressed all of these aspects over a period of four years and across two different domains of application – exploration of space and task / knowledge acquisition for manipulation. The present volume documents the results of the CoSy project. The CoSy project was funded by the European Commission as part of the Cognitive Systems Program within the 6th Framework Program.

Autoren/Hrsg.

Christensen, Henrik

Inhaltsverzeichnis

1;Preface;6
2;Contents;8
3;List of Contributors;10
4;Part I Introduction;13
4.1;Cognitive Systems Introduction;14
4.1.1;Introduction;14
4.1.2;Objective of Project;15
4.1.2.1;The Problem;15
4.1.2.2;The Way Forward;15
4.1.2.3;Steps to Success;15
4.1.3;A Motivating Example;18
4.1.4;Organization of the Research/Research Questions;20
4.1.4.1;Architecture;21
4.1.4.2;Representations;24
4.1.4.3;Learning;26
4.1.4.4;Perception-Action Modelling;28
4.1.4.5;Continuous Planning and Acting in Dynamic Multiagent Environments;29
4.1.4.6;Models of Action and Communication for Embodied Cognitive Agents;34
4.1.4.7;Multi-Modal Recognition and Categorisation;37
4.1.4.8;Scenario Driven Research;39
4.1.4.9;Exploration/Mapping of Space;40
4.1.4.10;Models for Object and Concepts;44
4.1.5;Consortium;50
4.1.6;Organization of the Book;51
4.1.7;References;52
5;Part II Component Science;60
5.1;Architecture and Representations;61
5.1.1;Introduction;61
5.1.2;Architectures and the Science of Cognitive Systems;62
5.1.3;Requirements for Architectures for Cognitive Robots;64
5.1.4;A New Architectural Schema;67
5.1.4.1;Key Features of CAS;67
5.1.4.2;Subarchitecture Design;68
5.1.4.3;System Wide Design;69
5.1.4.4;CAST: A Toolkit Implementing CAS;71
5.1.5;Four Problems;72
5.1.5.1;Binding;73
5.1.5.2;Filtering;82
5.1.5.3;Filtering Summary;86
5.1.5.4;Processing Management;86
5.1.6;The Relationship of CAS to Previous Work on Architectures;98
5.1.6.1;Cognitive Architectures;98
5.1.6.2;Robotic Architectures;99
5.1.7;Summary of Contributions and Conclusions;101
5.1.8;References;101
5.2;The Sensorimotor Approach in CoSy: The Example of Dimensionality Reduction;104
5.2.1;Introduction;104
5.2.2;Artificial Agents and Human Perception;105
5.2.3;Dimensionality Reduction;107
5.2.3.1;Review of Classical, “Passive” Approaches;107
5.2.3.2;Standard Issues;109
5.2.3.3;The Central Issue of the Metric;111
5.2.4;Dimension Reduction in the Context of Sensorimotor Interactions;114
5.2.4.1;Finding a Mathematical Framework;115
5.2.4.2;Back to Dimensionality Reduction;121
5.2.5;An Embodied Algorithm for Dimension Reduction;123
5.2.5.1;Description of the Algorithm;124
5.2.5.2;Results;129
5.2.5.3;Related Work;130
5.2.6;Conclusion;134
5.2.7;References;135
5.3;Categorical Perception;140
5.3.1;Introduction;140
5.3.1.1;Towards Hierarchical Scalable Representations;143
5.3.1.2;Towards Representations for Affordance-Based Categorization;143
5.3.1.3;Representations and Discovery of Object Classes by Generative Decompositions;144
5.3.1.4;Representations of Object Dynamics;144
5.3.2;Low-Level Features and Hierarchical Representation Learning;145
5.3.2.1;Towards Scalable Representations for Visual Categorization;145
5.3.2.2;Representations for Functional and Affordance-Based Categorization;152
5.3.3;Mid-Level Representation and Detection;157
5.3.3.1;Towards Adaptive Representations;157
5.3.3.2;Learning of Generative Decompositions;157
5.3.3.3;Generative/Discriminative Hybrid Model for Detection;160
5.3.3.4;Results on Visual Category Detection;161
5.3.3.5;Discussion;161
5.3.4;High-Level Representations and Dynamic Models;162
5.3.4.1;Appearance Model for Single-Frame Detection and Pose Estimation;163
5.3.4.2;Representing the Dynamics of the Human Walking Cycles with Latent Variable Model;164
5.3.4.3;Robust Detection and Tracking of People in Image Sequences;165
5.3.5;Outlook and Discussion;167
5.3.6;References;168
5.4;Semantic Modelling of Space;174
5.4.1;Introduction;174
5.4.1.1;Related Work;175
5.4.1.2;Outline;176
5.4.2;Background;176
5.4.3;Overview of the Spatial Model;177
5.4.3.1;Metric Map;177
5.4.3.2;Navigation Map;179
5.4.3.3;Topological Map;179
5.4.3.4;Conceptual Map;179
5.4.4;Metric Mapping;180
5.4.4.1;M-Space;180
5.4.4.2;Single Camera Bearing Only SLAM;183
5.4.4.3;Using Visual Attention for SLAM;184
5.4.4.4;Visual Scans;185
5.4.5;Navigation and Topological Maps;186
5.4.5.1;Building the Navigation Graph;187
5.4.5.2;Space Segmentation and Topological Graph;188
5.4.5.3;Adding Object Information;188
5.4.5.4;Adding Semantic Place Information;189
5.4.6;Conceptual Map;191
5.4.7;Object Detection and Recognition;193
5.4.7.1;Object Search and Localization;194
5.4.7.2;Object Distance Estimation;197
5.4.8;Place Classification;199
5.4.8.1;Multiple Cues and Modalities for Place Classification;201
5.4.8.2;Architecture of the Place Classification System;202
5.4.8.3;Laser-Based Place Classification;203
5.4.8.4;Vision-Based Place Classification;206
5.4.8.5;Discriminative Cue Integration;209
5.4.8.6;Adaptive Place Classification;211
5.4.9;Experiments with Place Classification;212
5.4.9.1;Single-Cue Place Classification;212
5.4.9.2;Combining Multiple Cues and Modalities;216
5.4.9.3;Adaptive Place Classification;218
5.4.9.4;Semantic Labeling of Space;219
5.4.10;Summary;222
5.4.11;References;224
5.5;Planning and Failure Detection;231
5.5.1;Introduction;231
5.5.2;The Multiagent Planning Language MAPL;234
5.5.3;Continual Planning;236
5.5.3.1;Assertions;236
5.5.3.2;Assertional Planning;238
5.5.4;Probabilistic Monitoring of Dynamic Processes;241
5.5.4.1;Sequential State Estimation;242
5.5.4.2;Particle Filters for Nonparametric Bayesian Filtering;244
5.5.4.3;Modeling the Influence of Failures Using Hybrid DBNs;245
5.5.5;Gaussian Processes Proposals for Failure Events;246
5.5.5.1;Data-Driven Proposal Distributions;247
5.5.5.2;Learning Sampling Models from Data;248
5.5.5.3;Predicting Collision Events and Parameters;250
5.5.6;Implementation of Sensor-Level Monitoring;251
5.5.6.1;Evaluation;252
5.5.7;Continual Collaborative Planning;253
5.5.8;MAPSIM;258
5.5.9;Situated Dialogue as Continual Collaborative Planning;260
5.5.10;Related Work;263
5.5.11;Conclusion;266
5.5.12;References;268
5.6;Multi-modal Learning;273
5.6.1;Introduction;273
5.6.2;Continuous Learning Framework;275
5.6.2.1;Introduction;275
5.6.2.2;Different Modes of Learning;278
5.6.2.3;Learning Algorithm;279
5.6.2.4;Reconstructive Representations for Interactive/Online Learning;281
5.6.2.5;Experimental Results;283
5.6.2.6;Discussion and Outlook;286
5.6.3;Cross-Modal Learning of Visual Categories;288
5.6.3.1;Object Representation by Scale-Invariant Patterns;289
5.6.3.2;Data-Driven Visual Grouping;290
5.6.3.3;Combining Unsupervised and Supervised Learning;291
5.6.3.4;Language System;292
5.6.3.5;Scene Reasoning;293
5.6.3.6;Label Propagation and Conflict Resolution;294
5.6.4;Learning Complex Actions;295
5.6.4.1;Introduction;295
5.6.4.2;Action Representation;296
5.6.4.3;Inference;299
5.6.4.4;Learning;301
5.6.4.5;Experiments;301
5.6.4.6;Conclusion;303
5.6.5;Functional Object Class Detection Based on Learned Affordance Cues;303
5.6.5.1;Related Work;304
5.6.5.2;Affordance Cue Acquisition;305
5.6.5.3;Functional Object Category Detection;306
5.6.5.4;Experiments;307
5.6.6;Conclusion and Outlook;308
5.6.7;References;313
5.7;Situated Dialogue Processing for Human-Robot Interaction;318
5.7.1;Introduction;318
5.7.2;Background;321
5.7.2.1;Multi-level Integration in Language Processing;322
5.7.2.2;Language Processing and Situational Experience;323
5.7.3;Talking;325
5.7.4;Talking about What You Can See;337
5.7.5;Talking about Places You Can Visit;348
5.7.5.1;Talking about Places;348
5.7.5.2;Representing Places to Talk about;349
5.7.5.3;Referring to Elsewhere;350
5.7.5.4;Understanding References to Elsewhere;354
5.7.6;Talking about Things You Can Do;356
5.7.7;Conclusions;358
5.7.8;References;360
5.7.9;Packing Algorithm;365
6;Part III Integration and Systems;372
6.1;The PlayMate System;373
6.1.1;Introduction;373
6.1.2;System Overview;379
6.1.2.1;Vision SA;381
6.1.2.2;Communication SA;382
6.1.2.3;Manipulation SA;383
6.1.2.4;Spatial SA;384
6.1.2.5;Binding;385
6.1.2.6;Motivation and Flow of Control;388
6.1.3;System Level Control of Information Flow;389
6.1.3.1;Cross Modal Learning;389
6.1.3.2;Clarification and Question Answering;393
6.1.3.3;Mediating between Qualitative and Quantitative Representations;396
6.1.4;Conclusions and Discussion;396
6.1.5;References;398
6.2;The Explorer System;400
6.2.1;Introduction;400
6.2.1.1;Related Work;402
6.2.1.2;Outline;402
6.2.2;System Overview;403
6.2.2.1;Navigation SA;403
6.2.2.2;Object SA;405
6.2.2.3;Place SA;405
6.2.2.4;Conceptual Mapping SA;406
6.2.2.5;The Robot Platforms;406
6.2.3;Spatial Modeling and Reasoning;407
6.2.3.1;Map Acquisition;407
6.2.3.2;Acquiring the Conceptual Map;409
6.2.3.3;Cross-Modal Spatial Knowledge Sharing;411
6.2.4;Planning;416
6.2.5;Scenario: Find Object;418
6.2.6;Conclusions;423
6.2.7;References;425
6.3;Lessons Learnt from Scenario-Based Integration;427
6.3.1;Introduction;427
6.3.2;But Is It Implemented?;428
6.3.3;Lessons;429
6.3.3.1;Integrate Ideas First;430
6.3.3.2;Integrate People Second;432
6.3.3.3;Choose Your Tools Wisely;433
6.3.3.4;Find the Scenario Sweet-Spot;436
6.3.3.5;Beware the Modularity Mantra;440
6.3.4;Conclusion;441
6.3.5;References;442
7;Part IV Summary and Outlook;443
7.1;Cross-Disciplinary Reflections: Philosophical Robotics;444
7.1.1;Introduction;444
7.1.2;Must an Intelligent Robot Use Language?;446
7.1.3;The Role of the Environment;447
7.1.4;Analysing Requirements Is Very Hard;448
7.1.5;Robotics and Philosophy of Science;449
7.1.5.1;Ontologies and Laws;449
7.1.5.2;No “Right” or “Best” Designs;450
7.1.5.3;A Science of Explosive Diversity;451
7.1.5.4;Individual Variability;452
7.1.5.5;The “Designer Stance” in Biology;452
7.1.5.6;Should Requirements Refer to Laws of Behaviour?;453
7.1.6;Environment-Neutral Requirements and Limitations;454
7.1.6.1;Forms of Representation;454
7.1.6.2;Architectures;456
7.1.7;De-fusing Diversity: Towers and Layers;457
7.1.7.1;Generative Frameworks;457
7.1.7.2;Subdivision into Towers of Functionality;458
7.1.7.3;Subdivision into Layers of Functionality;458
7.1.8;The CogAff Architecture Schema – One Small Step;460
7.1.8.1;Omega Architectures;461
7.1.9;Beyond the CogAff Schema;462
7.1.9.1;Where Are the Linguistic Mechanisms?;463
7.1.9.2;Varieties of Compositional Semantics;465
7.1.10;The H-CogAff Special Case;466
7.1.11;Study Trade-Offs Not Special Cases;467
7.1.11.1;Nature-Nurture Trade-Offs;468
7.1.11.2;Image-Scene Tradeoffs in Visual Processing;468
7.1.11.3;Trade-Offs Related to Noise and Uncertainty;469
7.1.12;Requirements for Visual Systems;470
7.1.12.1;Why Do Perception and Action Need Towers?;470
7.1.12.2;Multi-strand Process Perception;471
7.1.12.3;How to Acquire Useful Ontologies;472
7.1.12.4;Varieties of Complexity Reduction;472
7.1.12.5;Beyond J.J. Gibson’s Affordances;474
7.1.12.6;Implications of Speed of Human and Animal Visual Perception;475
7.1.13;Learning to Be a Mathematician;476
7.1.13.1;Two Kinds of Causation;477
7.1.14;Confusions about the Role of Embodiment;478
7.1.15;Developing the Revolution in Philosophy;479
7.1.16;Further Documentation on These Ideas;480
7.1.17;Why Other Disciplines Need AI;480
7.1.18;Conclusion: The Future;481
7.1.19;References;481
7.2;Lessons and Outlook;488
7.2.1;Introduction;488
7.2.2;Lessons;489
7.2.3;Outlook;491
8;Author Index;493