2. Learning in Artificial Sensorimotor Systems Daniel D. Lee

3. Synthetic intelligence u Hollywood versus reality Data HAL Robot Governor

4. Biological inspiration u Biological motivation for the Wright brothers in designing the airplane

5. Robot dogs u Platforms for testing sensorimotor machine learning algorithms Custom built versionSony Aibos

6. Robot hardware u Wide variety of sensors and actuators.

7. Robocup u Grand challenge for robotics community Legged league Humanoid league Middle size league Small size league Simulation league

8. Legged league u Recently implemented larger field and wireless communications among robots. u Each team consists of 4 Sony Aibo robot dogs (one is a designated goalie), with WiFi communications. u Field is 3 by 5 meters, with orange ball and specially colored markers. u Game played in two halves, each 10 minutes in duration. Teams change uniform color at half-time. u Human referees govern kick-off formations, holding, penalty area violations, goalie charging, etc. u Penalty kick shootout in case of ties in elimination round.

9. Upennalizers in action u GOOAAALLL! 2nd place in 2003.

10. Robot software architecture Cognition (Plan) Action (Actuators) Perception (Sensors) u Sense-Plan-Act cycle.

11. Perception View from Penn’s omnidirectional camera

12. Robot vision Color segmentation: estimate P(Y,Cb,Cr | ORANGE) from training images Region formation: run length encoding, union find algorithm Distance calibration: bounding box size and elevation angle Camera geometry: transformation from camera to body centered coordinates u Tracking objects in structured environment at 25 fps

13. Image reduction u 144  176  3 RGB image to 2 position coordinates Deterministic position: Probabilistic model (Kalman):

14. Image manifolds u Variation in pose and illumination give rise to low dimensional manifold structure Pixel vector

15. Learning nonlinear manifolds u Many recent algorithms for nonlinear manifolds. Kernel PCA, Isomap, LLE, Laplacian Eigenmaps, etc.

16. Locally linear embedding u LLE solves two quadratic optimizations using eigenvector methods (Roweis & Saul).

17. Translational invariance u Application of LLE for translational invariance.

18. Action M. Raibert’s hopping robot (1983)

19. Inverse kinematics u Degenerate solutions with many articulators.

20. Gaits u 4-legged animal gaits

21. Walking u Parameters tuned by optimization techniques Inverse kinematics to calculate joint angles in shoulder and knee

22. Behavior SRI’s Shakey (1970)

23. Probabilistic localization Particle filterKalman filter u Kalman and particle filters used to represent pose  x,y

24. Finite state machine u Event driven state machine. Search for ball Goto ball position Kick ball See ball Close to ball

25. Potential Fields u Charged particle dynamics to guide motion Potential fields for ball (attractive), field positions (attractive), robots (repulsive), penalty area (repulsive)

26. Learning behaviors u Reinforcement learning for control parameters Potential field parameters State selection Role switching Adaptive strategies

27. Reactive behaviors u “Behavior-based” robotics maps sensors to actuators without complex reasoning

28. Stimulus-response mapping Stimulus space Response space u Construct low dimensional representations for mapping stimulus to response

29. Image correspondences Object 1Object 2 u Correspondences between images of objects at same pose

30. Data from the web... u http://www.bushorchimp.com

31. Learning from examples Given Data (X 1,X 2 ): n labeled correspondences N 1 examples of object 1 (D 1 dimensions) N 2 examples of object 2 (D 2 dimensions) u Matrix formulation (n << N 1, N 2 ) ? ? D1N1D1N1 D1nD1n D2nD2nD2N2D2N2 X1X1 X2X2

32. Supervised learning u Problem overfitting with small amount of labeled data Fill in the blanks: (D 1 +D 2 )  n labeled data D 1  D 2 parameters ? ? D1N1D1N1 D1nD1n D2nD2nD2N2D2N2 Training Data

33. Supervised backprop network Original: Reconstruction: Original: Reconstruction: u 15 hidden units, tanh nonlinearity

34. Missing data u Treat as missing data problem using EM algorithm EM algorithm: Iteratively fills in missing data statistics, reestimates parameters for PCA, factor analysis ? ? D1N1D1N1 D1nD1n D2nD2nD2N2D2N2 D 1 +D 2

35. EM algorithm u Alternating minimization of least squares objective function. Y1Y1 Y2Y2 X Y so that X1X1 X2X2

36. PCA with correspondences u 15 dimensional subspace, 200 images of each object, 10 in correspondence Original:

37. Factor analysis Original: u 15 dimensional subspace with diagonal noise covariance

38. Common embedding space Two input spaces Common low dimensional space

39. LLE with correspondences u Quadratic optimization with constraints is solved with spectral decomposition Correspondences:

40. LLE with correspondences Original: u 8 nearest neighbors, 2 dimensional nonlinear manifold

41. Quantitative comparison u Incorporating manifold structure improves reconstruction error. Correspondence fraction Normalized error

42. Summary u Adaptation and learning in biological systems for sensorimotor processing. u Many sensory and motor activations are described by an underlying manifold structure. u Development of learning algorithms that can incorporate this low dimensional manifold structure. u Still much room for improvement…