SAIL Robot

1
SAIL Robot Developmental Robotics

• Resources
• Michigan State University Web site
• http//www.cse.msu.edu/weng/research/LM.html
• Papers
• J. Weng, J. McClelland, A. Pentland, O. Sporns,
  I. Stockman, M. Sur and E. Thelen, Autonomous
  Mental Development by Robots and Animals,''
  Science, vol. 291, no. 5504, pp. 599 - 600, Jan.
  26, 2000.
• http//www.cse.msu.edu/dl/SciencePaper.pdf
• J. Weng and Y. Zhang, Developmental Robots A
  New Paradigm,''  Invited paper in Proc. Second
  International Workshop on Epigenetic Robotics
  Modeling Cognitive Development in Robotic
  Systems.
• http//www.lucs.lu.se/ftp/pub/LUCS_Studies/LUCS94/
  Weng.pdf
• J. Weng and W. Hwang, An incremental learning
  algorithm with automatically derived
  discriminating features'',  in Proc. Asian
  Conference on Computer Vision,  Taipei, Taiwan,
  pp. 426 - 431, Jan. 8 - 9, 2000.
• http//www.cse.msu.edu/weng/research/ACCV2000.pdf
  
• Video
• See MSU web site

2
Developmental Robotics

• A.k.a. Epigenetic Robotics
• Principles from biological psychological
  development
• Development as long-term learning
• A prolonged developmental process through which
  varied and complex cognitive and perceptual
  structures emerge as a result of an embodied
  system interacting with a physical and social
  environment.
• Development relies on
• Some innate components
• An environment
• Interaction between innate components
  environment
• E.g., language development (Elman, Bates,
  Johnson, Karmiloff-Smith, Parisi, Plunkett,
  1996)

3
What is developmental?
world4
biology
conception
1
2
3
4
5
Time (years)
4
Why?

• Formal models for developmental psychology
• Often psychological models are informal
• E.g., Hollich, Hirsh-Pasek, Golinkoff (2000)
• Formal models, including robots, can help us
  think about child development
• Engineering rationales
• Enabling robots and other artificial systems to
  better adapt to their environments, and to better
  adapt to changes in these environments
• Benefits for Human Computer Interaction (HCI)
• Easing the problem of programming robotsby
  programming the robots to develop skills for any
  particular environment instead of programming
  robots for specific environments (see also Weng,
  McClelland, Pentland, Sporns, Stockman, Sur,
  Thelen, 2001).

5

• combining a limited amount of a priori
  knowledge with learning has not been
  extensively explored in robotics , especially
  in the context of nonlinear redundant systems
  (Berthouze Kuniyoshi, 1998)

6
SAIL Robot

• IHDR algorithm
• Incremental Hierarchical Discriminant Regression
• Generates a mapping from high-dimensional input
  vectors to high-dimensional output vectors
• Input sensory input state ( motor guidance)
• Output new state motor output
• Mapping maps sensory input state to output
• A human trainer shapes behaviors by providing
  motor guidance
• Motor commands then associated with the current
  sensory input state
• Upon repeated use of the same motor commands by
  the human trainer, the features of the sensory
  input state that discriminate use of this motor
  command are learned
• Example
• Trainer pushes robot to left, when approaching a
  left turn in corridor
• Sensory stimulus and internal state relevant to a
  left turn are discriminated and associated with
  the motor outputs to generate a left turn
• Features unrelated to the left turn are
  disregarded (e.g., a person walking by the robot
  in the hall)

7
IHDR (Incremental Hierarchical Discriminant
Regression)

• Decision tree representation
• applications of decision trees have been
  traditionally for a low-dimensional feature space
  with manually selected features (p. 426, Weng
  Hwang, 2000)
• Sensors and effectors can be high-dimensional
• E.g., a 640x480 resolution camera 307, 200
  length vector of pixels
• IHDR automatically derives discriminating
  features
• Decision trees we have looked at have features
  given
• Method
• Goal Approximate a mapping (function)
• H X ? Y, given a set of training samples
• (xi, yi) xi ? X, yi ? Y, i1, 2, n
• X is sensory input internal state of robot
• Y is motor output
• Creates a decision tree (HDR tree) with nodes
  containing
• y-clusters, and x-clusters
• Each node has at most q (e.g., q6) clusters of
  each type
• q is branching factor of tree

8
Training Testing

• This is a supervised method
• Training
• Input
• A new (x, y) pair
• That is, both the sensory state input (x) and
  correct output must be given
• Processing
• Adding (x, y) to existing node, or building a new
  node in tree
• y is used to find node with the the nearest
  y-cluster (using Euclidean distance)
• The x value will be in the x-cluster for this
  node, or a subnode
• x used to compute statistics (mean covariance
  matrix)
• Compute probability that (x, y) belongs to
  existing cluster in the tree
• Output
• Change to the tree
• Testing
• Input
• An x value sensory state input
• Processing
• x used to compute statistics (mean covariance
  matrix)
• Compute probability that (x, y) belongs to
  existing cluster in the tree

9
Up to q clusters
Up to q children per node

Etc.
10
Additional References

• Berthouze, L. Kuniyoshi, Y. (1998). Emergence
  and categorization of coordinated visual behavior
  through embodied interaction. Machine Learning,
  31.
• Elman, J. L., Bates, E. A., Johnson, M. H.,
  Karmiloff-Smith, A., Parisi, D., Plunkett, K.
  (1996). Rethinking Innateness A Connectionist
  Perspective on Development. Cambridge, MA MIT
  Press.
• Hollich, G. J., Hirsh-Pasek, K., Golinkoff, R.
  M. (2000). Breaking the language barrier An
  emergentist coalition model for the origins of
  language learning. Monographs of the Society for
  Research in Child Development, Serial No. 262,
  Vol. 65, No. 3.
• Dan Swets and John J. Weng, Using Discriminant
  Eigenfeatures for Image Retrieval,'' IEEE Trans.
  on Pattern Analysis and Machine Intelligence,
  vol. 18, no. 8, pp. 831-836, Aug. 1996.