Learning in Artificial Sensorimotor Systems Daniel D. Lee.

Learning in ArtificialSensorimotor Systems

Daniel D. Lee

Synthetic intelligence

Hollywood versus reality

Data

HAL

Robot Governor

Biological inspiration

Biological motivation for the Wright brothers in designing the airplane

Robot dogs

Platforms for testing sensorimotor machine learning algorithms

Custom built version Sony Aibos

Robot hardware

Wide variety of sensors and actuators.

Robocup

Grand challenge for robotics community

Legged league Humanoid league

Middle size leagueSmall size league

Simulation league

Legged league

Recently implemented larger field and wireless communications among robots.

Each team consists of 4 Sony Aibo robot dogs (one is a designated goalie), with WiFi communications.

Field is 3 by 5 meters, with orange ball and specially colored markers.

Game played in two halves, each 10 minutes in duration. Teams change uniform color at half-time.

Human referees govern kick-off formations, holding, penalty area violations, goalie charging, etc.

Penalty kick shootout in case of ties in elimination round.

Upennalizers in action

GOOAAALLL! 2nd place in 2003.

Robot software architecture

Cognition(Plan)

Action(Actuators)

Perception(Sensors)

Sense-Plan-Act cycle.

Perception

View from Penn’somnidirectional camera

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

Tracking objects in structured environment at 25 fps

Image reduction

144 176 3 RGB image to 2 position coordinates

€

(53,65,27) (52,67,35) L

(48,68,31) O

M (250,213,196)

⎡

⎣

⎢ ⎢ ⎢

⎤

⎦

⎥ ⎥ ⎥

€

(xball , yball )

€

P(xball , yball )

Deterministic position:

Probabilistic model (Kalman):

Image manifolds

Variation in pose and illumination give rise to low dimensional manifold structure

Pixel vector

€

0.0

0.5

0.7

1.0

M

0.8

0.2

0.0

⎡

⎣

⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢

⎤

⎦

⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥

Learning nonlinear manifolds

Many recent algorithms for nonlinear manifolds.

Kernel PCA, Isomap, LLE, Laplacian Eigenmaps, etc.

Locally linear embedding

LLE solves two quadratic optimizations using eigenvector methods (Roweis & Saul).

€

E(W ) =r X i − Wij

r X j

j∑

2

i∑

€

Φ(Y ) =r Y i − Wij

r Y j

j∑

2

i∑

Translational invariance

Application of LLE for translational invariance.

Action

M. Raibert’s hoppingrobot (1983)

Inverse kinematics

QuickTime™ and aTIFF (Uncompressed) decompressor

are needed to see this picture.

Degenerate solutions with many articulators.

€

H = R(θ1)oT (l1)oR(θ2 )oT (l2 )oR(θ3)oT (l3)

€

θ1

€

θ2

€

θ3

€

l1

€

l2

€

l3

Gaits

4-legged animal gaits

Walking

Parameters tuned by optimization techniques

Inverse kinematics to calculate joint angles in shoulder and knee

Behavior

SRI’s Shakey (1970)

Probabilistic localization

∑==

iii yxPP

PyxP

yxP

)|,()(

)()|,(

),,(

θθ

θθ

θ

Particle filter Kalman filter

Kalman and particle filters used to represent pose

x,y

Finite state machine

Event driven state machine.

Search forball

Goto ballposition

Kick ball

Seeball

Closeto ball

Potential Fields

Charged particle dynamics to guide motion

Potential fields forball (attractive), field positions

(attractive), robots (repulsive), penalty

area (repulsive)

Learning behaviors

Reinforcement learning for control parameters

AttackSupport

Defend

Goalie

Potential field parametersState selectionRole switchingAdaptive strategies

Reactive behaviors

“Behavior-based” robotics maps sensors to actuators without complex reasoning

Stimulus-response mapping

Stimulus space Response space

Construct low dimensional representations for mapping stimulus to response

Image correspondencesObject 1 Object 2

Correspondences between images of objects at same pose

Data from the web...

http://www.bushorchimp.com

Learning from examplesGiven Data (X1,X2):

n labeledcorrespondences

N1 examplesof object 1

(D1 dimensions)

N2 examplesof object 2

(D2 dimensions)

Matrix formulation (n << N1, N2 )

?

?D1N1D1n

D2n

D2N

2

X1

X2

Supervised learning

Problem overfitting with small amount of labeled data

Fill in the blanks:

(D1 +D2 ) n labeled dataD1 D2 parameters

?

?D1N1D1n

D2n

D2N

2

TrainingData

Supervised backprop network

Original:

Reconstruction:

Original:

Reconstruction:

15 hidden units, tanh nonlinearity

Missing data

Treat as missing data problem using EM algorithm

EM algorithm:

Iteratively fills in missing data statistics, reestimatesparameters for PCA, factor analysis

?

?D1N1D1n

D2n

D2N

2

D1+D2

EM algorithm

€

minYic=Yjc

minWi

Xi −WiYii

∑ 2

€

Yi ← RYi

Alternating minimization of least squares objective function.

€

Y TY = I Y1

Y2

X

Yso that

€

Wi ←WiR−1

X1

X2

PCA with correspondences

15 dimensional subspace, 200 images of each object, 10 in correspondence

Original:

Original:

Factor analysis

Original:

Original:

15 dimensional subspace with diagonal noise covariance

Common embedding space

Two input spaces Common low dimensional space

LLE with correspondences

Quadratic optimization with constraints is solved with spectral decomposition

Correspondences:

€

E(W1) =r X i

1 − Wij1 r X j

1

j∑

2

i∑

€

E(W 2 ) =r X i

2 − Wij2 r X j

2

j∑

2

i∑

€

Φ(Y1,Y 2 ) =r Y i

1 − Wij1 r Y j

1

j∑

2

i∑

€

+r Y i

2 − Wij2 r Y j

2

j∑

2

i∑

€

i∈Sc :r Y i

1 =r Y i

2

LLE with correspondences

Original:

Original:

8 nearest neighbors, 2 dimensional nonlinear manifold

Quantitative comparison

Incorporating manifold structure improves reconstruction error.

Correspondence fraction

Nor

mal

ized

err

or

Summary

Adaptation and learning in biological systems for sensorimotor processing.

Many sensory and motor activations are described by an underlying manifold structure.

Development of learning algorithms that can incorporate this low dimensional manifold structure.

Still much room for improvement…