CS 395T: Intelligent Robotics

• Spring 2006. TTh, 3:30 - 5:00 pm
• TAY 3.144
• Unique #54215.
• Professor Benjamin Kuipers
• Office hours: 1-2 pm, TTh.

Syllabus

Traditionally, robots have been useful in manufacturing by moving blindly but precisely in totally controlled workcells. Traditionally, symbolic AI systems have given the appearance of intelligence by applying logical inference algorithms to symbol structures whose primitive elements are specified by human programmers. This has left AI systems open to Searle's famous "Chinese Room" critique, arguing that they only mimic intelligence: they are merely "faking it".

To answer this philosophical challenge, and to be useful in a host of real-world application on Earth and in space, AI systems need to be robots, with sensors and effectors embedded in the physical world. Not only that, but these robots must learn the nature of their own sensorimotor interaction with the environment, and must create their own symbols, grounded in their own experience.

Robots are being created with ever more complex and richly structured sensors. The sensorimotor system evolves over time, sometimes deteriorating, but sometimes being augmented with new "plug-and-play" sensors. Humans are astonishingly adaptable to sensorimotor changes, and children do an amazing job of learning to use their sensors and effectors in a few short years after birth. We can learn important things about robots from research on children. And robot models may help us create better theories of child development.

Our main focus is on robot learning of the "foundational domains" that underlie commonsense knowledge: space, time, actions, objects, properties and affordances, causality, and so on. The question is how these higher-level ontologies can be learned from experience with low-level sensor and motor interactions with the world (the "pixel level").

Perhaps surprisingly, this has led us to certain questions about the nature of consciousness, what its properties are, what those properties tell us about the architecture of the mind, and how it is involved in learning these foundational domains. We will read books and papers from a wide range of disciplines, attempting to see the underlying reality that they are describing from different points of view.

Assignments

The requirements of the course will be:

• (35%) Class presentation on one or more papers.
• (15%) Class participation in discussions.
• (50%) Term project, presentation, and paper.

Class Presentations

Pick a presentation topic that works well with your term project topic. The papers will be accessible online through the UT Library, or via link here. In some cases, you will need to review several related papers by the authors.

Here is a thematic outline. You don't need to cover the points in exactly this order, but try to address these needs for your audience.

• What is the problem? Why is it important? Why should the reader care?
• What assumptions are being made?
• How does this method work?
  Provide an intuition to guide the hearer through the technical details.
  Then provide a more detailed example to show what the intuitions mean.
• What are the strengths and limitations of this approach?
• How can you evaluate the benefits?
• What are open problems in this area?

• How does this help us?
  Where is the gold?

Presentation Topics

• Introduction (2 weeks: Jan 17 - 26) [Kuipers]
  • The Hybrid Spatial Semantic Hierarchy [Beeson and Modayil]
  • Progress in Foundational Learning [Provost and Modayil]
  • Course intro: How can a robot have a mind of its own?
  • Learning about an uninterpreted sensorimotor system
    Read [Pierce and Kuipers, AIJ, 1997], which is based on our Spatial Semantic Hierarchy work on cognitive mapping in [Kuipers, AIJ, 2000].

• Statistical Learning Methods (2 weeks: Jan 31 - Feb 9)
  • Clustering: AutoClass
    • [Cheeseman & Stutz, 1996] [.ps]
  • Factoring: Principal Component Analysis (PCA) and Independent Component Analysis (ICA)
    • Hyvarinen, Survey on independent component analysis.
      Neural Computing Surveys 2: 94-128, 1999. [PDF]
      
    • Steyvers, Multidimensional Scaling.
      Encyclopedia of Cognitive Science, 2002. [PDF]
      
  • Clustering: Self-Organizing Maps and Growing Neural Gas
    • Concise description of SOMs, plus a huge bibliography.
      
    • Fritzke, A growing neural gas network learns topologies.
      NIPS 1994. [PDF]
      
  • Factoring: Nonlinear dimensionality reduction
    • Tenenbaum, de Silva and Langford, A global geometric framework for nonlinear dimensionality reduction. Science 290: 2319-2323, 2000. [PDF]
    • Roweis and Saul, Nonlinear dimensionality reduction by locally linear embedding. Science 290: 2323-2326, 2000. [PDF]

• Learning about Actions (2 weeks: Feb 14 - 23)
  • Marginal attribution: Drescher's schema mechanism [Kuipers]
    • Gary Drescher, Made-Up Minds, MIT Press, 1991. [Amazon]
  • Bayesian structure learning in children
    • Sobel, Tenenbaum and Gopnik, Children's causal inferences from indirect evidence. Cognitive Science, 2004. [PDF]
    • See related work by Josh Tenenbaum.
  • Learning behavior models from observations
    • Fox, Ghallab, Infantes, and Long, Robot introspection through learned hidden Markov models. Artificial Intelligence 170: 59-113, 2006. [PDF]
  • Curiosity-driven exploration (Oudeyer and Kaplan)
    • Oudeyer and Kaplan, Intelligent adaptive curiosity. Epigenetic Robotics Conference, 2004. [PDF]

• Computer and Human Vision (2 weeks: Feb 28 - Mar 9)
  • Vision as an embodied process
    • Ballard, Hayhoe, Pook, and Rao, Deictic codes for the embodiment of cognition. Behavioral and Brain Sciences 20: 723-767, 1997. [PDF]
  • Vision as mastery of sensorimotor contingencies
    • O'Regan and Noe, A sensorimotor account of vision and visual consciousness. Behavioral and Brain Sciences 24: 939-1031, 2001. [PDF]
  • Recognizing and learning object categories
    • ICCV-05 short course by Fei-Fei, Fergus, and Torralba.
  • Tracking visual objects
    • Forsythe and Ponce, Computer Vision: A Modern Approach, 2003, ch.17 [Amazon] (but note reviews)
    • Peter Corke, The Machine Vision Toolbox, IEEE Robotics and Automation Magazine, December 2005. [PDF]

• Connecting Symbols to the World (3 weeks: March 21 - April 6)
  • Semiotic schemas
    • Deb Roy, Semiotic schemas: a framework for grounding language in action and perception. Artificial Intelligence 167: 170-205, 2005. [PDF]
  • Learning to talk about events
    • Dominey and Boucher, Learning to talk about events from narrated video in a construction grammar framework. Artificial Intelligence 167: 31--61, 2005. [PDF]
  • Evolving language and grammar
    • Paul Vogt, The emergence of compositional structure in perceptually grounded language games. Artificial Intelligence 167: 206-242, 2005. [PDF]
  • Embodied multimodal language learning
    • Yu and Ballard, On the integration of grounding language and learning objects. AAAI, 2004. [PDF]
    • Yu and Ballard, A multimodal learning interface for grounding spoken language in sensorimotor experience. ACM Trans. Applied Perception 1, 2004. [PDF]
    • Also see Yu, Ballard and Aslin, The role of embodied intention in early lexical acquisition. Cognitive Science 29(6): 961-1005, 2005. [PDF]
  • Image schemas (Mark Johnson; Beate Hampe)
    • Mark Johnson, The Body in the Mind, 1987. [Amazon]
  • TBD

• Cognitive Architecture (2 weeks: April 11 - 20) [Kuipers]
  • Consciousness: responding to John Searle
    • Searle, Mind: A Brief Introduction, 2004, chap.5. [Amazon]
    • Kuipers, Consciousness: Drinking from the firehose of experience, AAAI-05. [PDF]
  • Global Workspace Theory according to Bernard Baars
    • Baars, A Cognitive Theory of Consciousness, 1988. [online] [Amazon]
    • Baars, In the Theater of Consciousness, 1997. [Amazon]
  • Perceptual Meaning Analysis according to Jean Mandler
    • Mandler, The Foundations of Mind, 2004. [Amazon]
  • TBD

• Project Reports (2 weeks: April 25 - May 4)

Term Projects

You are encouraged to select a topic that fits well with your other research interests.

Possible Project Topics

• Distinguishing and characterizing sensor modalities. Can we use modern Bayesian methods to do a better job than [P+K,97] of separating an uninterpreted sense vector into distinct sensor modalities? As an example, here is a directory of sensor traces collected from Lassie, a Magellan Pro robot with laser range-finder, sonar, infrared, bump sensors, and odometry. Once these are separated into modalities, can you also distinguish between useful values and error values of each sensor? Can you find additional sensor traces on the web for testing your method? Also, look at some interesting related work by David Philipona on uninterpreted sensorimotor systems.

• Identifying cross-modal relationships. Given a separation of the sense vector into distinct modalities, replicate the multidimensional scaling of each modality from [P+K,97]. Can you determine how the different sensory modalities relate to each other, providing different kinds of information about the same underlying reality. For example, laser and sonar are both range sensors, but they have different units, different resolution, different coverage, different error models, and so on. IR and bump sensors are also range sensors, but with even more differences. Odometry is related to range, but less directly. Can you detect and characterize that relationship? Can you use the cross-modal relationships to find new concepts? Look at the work of Michael Coen on cross-modal clustering (AAAI-05) and multi-modal integration (IJCAI-01). (His recent work is on clustering of speech sounds and bird songs. Can we apply it to robot sensors?)

• Learning allocentric space from egocentric experience. In an egocentric frame of reference, when the robot acts, everything in its world moves around it. In a world-centered ("allocentric") frame of reference, the static parts of the world have a fixed location and only the robot itself moves (for now, assume the robot is the only dynamic part of the environment). A range-sensor trace captures the robot's egocentric experience with its environment. Is there an analysis method (clever use of ICA?) that will identify a useful world-centered ("allocentric") frame of reference for that environment? This amounts to creating a map of the static world, and updating the robot's pose as it moves within that frame of reference.

• Learning sensors with overlapping receptive fields. Vulcan, our intelligent wheelchair, has a pair of laser rangefinders mounted at its front corners, facing 45 degrees outward from the front axis, so their fields of view overlap by about 90 degrees. Assume (or demonstrate) that you can determine that these are two different sensors of the same type. Can the robot autonomously learn that the fields of view overlap? How should it represent the sensor input in the overlap region? How should it represent the sensor input in the non-overlapping regions? Can it build a better fused image of its nearby surround, using the information from the two laser rangefinders?

• Learning the structure of the retina. Vision provides vastly more information than laser range-finders, and vision is foveated, with much higher resolution in the fovea than in the periphery. Can the structure of the retina be learned from experience the way we have described for range-finders? (This learning could take place during evolution or embryogenesis rather than infancy.) For information on child development of vision, see Hainline, The development of basic visual abilities, in Slater (Ed.), Perceptual Development, 1998 [PDF]. (Also, be sure to read the O'Regan and Noe paper in connection with this project.)

  A simple way to define an experimental foveated retina is to create a version of the "roving eye" consisting of three concentric square retinas. The center one has full resolution of the image it moves over. The next one out has half the resolution, so its "pixels" represent 2x2 regions of the image. The outer one is half again, so its "pixels" are 4x4 regions in the image.

• Learning binocular tracking with foveated retinas. Suppose an agent has two eyes with foveated retinas (forward-facing like humans). The correlation between the images in the two eyes will be maximized when the two eyes have the same object in the foveal region (because that's where most of the pixels are). Can an agent that starts without the ability of visual tracking, learn that capability by learning to move the two eyes to keep their correlation maximized? This would allow tracking to be learned before object recognition, and would provide more information for learning object recognition later. If you can come up with a learning scheme for this, how does it relate to the evidence about how human infants learn visual tracking? (Note that there are many online databases of test images. Google: stereo pair image database)

• Learning object-based stereo range-finding from foveated tracking. Suppose an agent has a stereo pair of foveated eyes, and the capability of tracking maximally correlated image-pairs. Even looking at a single stereo pair of images, there will be multiple local correlation maxima, corresponding to objects at different distances and directions from the agent. Can the agent learn to use stereo disparity, the difference in axis-offset of the two retinas, to estimate the distance to the object? Over what range of distances would this be effective? What would be required for the agent to learn that this feature represents distance?

• Learning primitive actions for physical robots. Given sensor traces separated into distinct sensor modalities, replicate the multidimensional scaling from [P+K,97] that gives the layout of the sensor array. Then replicate the identification of primitive actions from [P+K,97] by collecting the average sensory flow vectors that result from random unit motor vectors. Perform PCA on these high-dimensional vectors to determine the natural motor primitives for the physical robot. Apply this to several different physical robots.

• Determine the layout of general distributed sensor networks. Many network experts are developing methods to determine the embedding geometry of distributed networks of various kinds, including our own Professor Lili Qiu. Can we make progress on this problem, using our methods for determining sensor array geometry basing inter-pixel distance estimates on the similarity of the sense information provided? Review the literature on sensor network geometry, discuss the issue with Lili Qiu and her students, think deeply about the question, and show what our methods can contribute to this problem.

• Autonomous learning of sensory features, distinctive states, and high-level actions. Jefferson Provost's SODA uses self-organizing maps to learn distinctive sensory features, then learns hill-climbing control laws to maximize the features at distinctive states, and trajectory-following control laws to move from one distinctive state to the neighborhood of the next. His work is done using a simulated robot. Can you replicate it using a sensor trace from Lassie? Use the available sensor trace (repeatedly) to train the SOM. Use the same sensor trace to build an occupancy grid map of the environment it was exploring. Then use that map in Player/Stage as the environment for further exploration by a simulated robot with the same sensor. Use the simulator to get the large amounts of experience required to train the hill-climber and trajectory-follower. Then show that the learned high-level actions can be used to control Lassie's physical travel in the environment. (This project is very ambitious. If you take it on, we may have to collect new sensor traces for you, but you can start with the existing ones.)

• Prerequisites and consequences of reliable actions. Gary Drescher's schema mechanism can be seen as a method for searching for the prerequisites and consequences of actions, evaluating schemas for their reliability. In Harold Chaput's PhD thesis, he used hierarchical SOMs to create a much more tractable replication of Drescher's schema mechanism. Understand and replicate Chaput's Constructivist Learning Architecture (CLA). Devise a new simulated-robot learning task, to evaluate CLA as a way to learn reliable higher-level actions. Does it complement or compete with SODA?

• Representing directly-perceived quantities. In very low-level learning and cognition, we face the foundational problem of how to represent basic directly-perceived quantities, ranging from distance and angle, to color or texture, to the magnitude of a sensation such as weight or loudness. What is stored? How can it be retrieved?

  One hypothesis is that what is stored is a high-resolution representation of the actual sensory input. However, this cannot be used directly to generate a response to a query. Rather, a proposed response is generated (physically or virtually), starting at some default value. Only an ordinal comparison (greater, equal, less) is possible between the stored value and the proposed response. The proposed response is adjusted until the two "match" (within reasonable tolerances). This hypothesis appears to match certain psychophysical experiments. Review the psychophysics literature on this, create and implement a computational model suited to the experimental tasks in the literature, and test your model against the observed results. Does the model make new predictions that could be tested?

Textbooks

The following books are valuable reading for the course.

• Bernard J. Baars. In the Theater of Consciousness. Oxford University Press, 1997.
  Available as a required book at the University Co-op.

• Jean Mandler. The Foundations of Mind: Origins of Conceptual Thought. Oxford University Press, 2004.
  Quite important; I recommend that you order a copy.

• Mark Johnson. The Body in the Mind: The Bodily Basis of Meaning, Imagination, and Reason. University of Chicago Press, 1987.

• Gary L. Drescher. Made-Up Minds: A Constructivist Approach to Artificial Intelligence. MIT Press, 1991.

• George Lakoff and Mark Johnson. Metaphors We Live By, second edition. University of Chicago Press, 2003.

Valuable books for your library

• Duda, Hart and Stork. 2001. Pattern Classification, Second Edition. NY: John Wiley and Sons.

• Tom Mitchell. 1997. Machine Learning. Boston: McGraw-Hill.
  (This book is a useful reference, and is the required text for Ray Mooney's Machine Learning course.)

• Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice-Hall.

The Computer Science Department has a Code of Conduct that describes the obligations of faculty and students. Read it at http://www.cs.utexas.edu/users/ear/CodeOfConduct.html.