Affordance-based Perception, Learning and Planning using Range - - PowerPoint PPT Presentation

▶
affordance based perception learning and planning using
SMART_READER_LITE
LIVE PREVIEW

Affordance-based Perception, Learning and Planning using Range - - PowerPoint PPT Presentation

Nov 27, 2023 226 likes •697 views

Affordance-based Perception, Learning and Planning using Range Images Erol ahin KOVAN Research Lab. Dept of Computer Eng. Middle East Technical University Ankara, TURKEY http://kovan.ceng.metu.edu.tr Dept. of Computer Engineering 1

slide-1
SLIDE 1

1

  • Dept. of Computer Engineering

Middle East Technical University

Affordance-based Perception, Learning and Planning using Range Images

Erol Şahin

KOVAN Research Lab. Dept of Computer Eng. Middle East Technical University Ankara, TURKEY http://kovan.ceng.metu.edu.tr

slide-2
SLIDE 2

2

Behavior development in psychology

  • J. Piaget & E.J. Gibson

Reflexes Repeated interaction Controlled action Discovery of capabilities

Developmental/Epigenetic Robotics

Simple pre-coded behaviors Interact with the environment Discover general relations Use relations in goal-directed behavior Affordances

M.Cakmak, M.R.Dogar, E. Ugur and E.Sahin. Affordances as a Framework for Robot Control. Proc. of EpiRob’07.

slide-3
SLIDE 3

3

Learning of Affordances

  • Differentiation: discovering distinctive features

and invariant properties in the environment

  • Exploratory activities that bring about changes

in the environment that an action produces

  • Perceptual Learning: develop an anticipation of
  • utcomes based on perception of invariants,

actions become performatory

E.J. Gibson (1910–2002)

slide-4
SLIDE 4

4

Equivalence Classes

Entity Equivalence Behavior Equivalence Affordance Equivalence (effect, (<entity>, behavior)) (effect, (entity, <behavior>)) (effect, <(entity, behavior)>)

Effect Equivalence

slide-5
SLIDE 5

5

Experimental Framework (1/2)

  • 6 wheel, differential drive
  • SICK laser range finder
  • 3-D scanning

– 0.25° horizontal resolution (180°) – 0.23° vertical resolution (165°)

  • MACSim : High-fidelity simulation

environment

  • ODE used as physics engine
  • Sensors and actuators are

calibrated

slide-6
SLIDE 6

6

Experimental Framework (2/2)

Primitive behaviors Limited motor capability Pre-coded

slide-7
SLIDE 7

7

Perceptual Features

More than 30000 perceptual features!

slide-8
SLIDE 8

8

Part 1: Perceptual Learning of Affordances

slide-9
SLIDE 9

9

Traversability for KURT3D

slide-10
SLIDE 10

10

Representation of entity and effect

Features(initial) Success/Fail entity

MOVE_FORWARD

behavior effect 3000 x Perceive Act Get result

A single interaction

  • E.Ugur, M.R.Dogar, M.Cakmak and E.Sahin. The Learning and Use of Traversability Affordances using

Range Images on a mobile robot. ICRA, Rome, Italy, April 2007.

slide-11
SLIDE 11

11

Learning: Selecting relevant features

Feature Selection (ReliefF) . . .

Entity Effect: fail Entity Effect: success Entity Effect: fail

. . .

Filtered Entity Filtered Entity Effect: fail Filtered Entity Effect: fail Effect: success

3000 interactions

slide-12
SLIDE 12

12

Scan only this region

Perceptual Economy

Only 1% of the features are relevant!

slide-13
SLIDE 13

13

. . .

Filtered Entity Effect: fail Filtered Entity Effect: fail Filtered Entity Effect: success

SVM Target values input

Learning: Mapping entity to effect

3000 interactions

slide-14
SLIDE 14

14

Prediction of Traversibility for Novel Objects

TRAIN 100% 100% 83.8% TEST 86% 100%

prediction prediction Successful generalization over novel objects

slide-15
SLIDE 15

15

Traversability in a cluttered environment

slide-16
SLIDE 16

16

Traversibility on KURT3D

slide-17
SLIDE 17

17

Pass-thru-ability on KURT3D

  • The training used only single
  • bject interactions.
  • The robot has not concept of an
  • bject or a gap.
  • The robot does not have any idea
  • n the size of its body.
  • Yet, one can see an affordance

ratio here..

How would Warren et al. comment?

slide-18
SLIDE 18

18

Curiosity-Driven Learning

  • Large number of training samples were required:

– ~3000 virtual interactions with environment. – Learning process is costly, time-consuming, risky.

  • Minimize number of interactions with minimal

degradation in learning process. 2 phase learning:

– Bootstrapping: small number of interactions

  • Learn the relevant features
  • Initiate an SVM model.

– Curiosity-driven

  • Interact with the environment and update SVM only if the

current situation is an interesting one

  • E.Ugur, M.R.Dogar, M.Cakmak and E.Sahin. Curiosity-driven Learning of Traversability Affordance on a

Mobile Robot. ICDL, London, UK, July 2007.

slide-19
SLIDE 19

19

Not interesting!

Probably traversable No object in the vicinity Probably non-traversable Cylinder object is very close.

slide-20
SLIDE 20

20

Interesting

Cylinder’s surface is similar to sphere’s Object is located at the boundaries for Go-forward action

slide-21
SLIDE 21

21

Curiosity-Driven Learning

PERCEPT SVM (trained upto now)

  • Within curiosity band
  • Classifier is less certain about hypothesized effect
  • Execute behavior
  • Learn from this sample. (Update SVM).
  • Outside curiosity band
  • Classifier is more certain about hypothesized effect
  • Do not execute behavior
  • Skip this sample.
slide-22
SLIDE 22

22

Two parameters of learning

  • Bootstrapping duration: No improvement after a certain level (25)
  • Width of curiosity-band is optimized. (0.5)
  • With bootstrapping of 25 and curiosity of 0.5, 200 interactions

delivers as good/better performance than 3000 interactions!

slide-23
SLIDE 23

23

Performance comparison

Curiosity-based interaction is more economical

slide-24
SLIDE 24

24

Part 2: From Primitive to Goal-Directed Behaviors

slide-25
SLIDE 25

25

Representation of entity and effect

Features(initial) Features(final) entity

MOVE_FORWARD

behavior effect Features(final-initial) 3000 x Perceive Act Perceive

A single interaction

M.R.Dogar, M. Cakmak, E. Ugur and E. Sahin. From Primitive Behaviors to Goal-Directed Behavior Using Affordances. IROS, San Diego, October 2007.

slide-26
SLIDE 26

26

Learning: Cluster Effects

. . .

Effect prototype-1

mean mean mean

Effect prototype-2 Effect prototype-10

. . . 3000 interactions

Entity Effect Entity Effect Entity Effect

For each behavior: K-means (k=10)

. .

Effect category-1 Effect category-2 Effect category-10

No fail/success criteria!

slide-27
SLIDE 27

27

Learning: Selecting relevant features

Feature Selection (ReliefF) . . . 3000 interactions

Entity Effect category Entity Effect category Entity Effect category

. . .

Filtered Entity Effect category Filtered Entity Effect category Filtered Entity Effect category

slide-28
SLIDE 28

28

Learning: Mapping entity to effect

. . . 3000 interactions

Filtered Entity Effect category Filtered Entity Effect category Filtered Entity Effect category

SVM Target values input

slide-29
SLIDE 29

29

Goal-directed Behaviors

Perception

Behavior Selection

…

Execution Perception

Behavior Selection

slide-30
SLIDE 30

30

Goal: Approach

slide-31
SLIDE 31

31

Goal-directed Behaviors on KURT3D

Goal is provided at run-time not during learning!

slide-32
SLIDE 32

32

Blending Behaviors

Similarity 0.4 x 0 x 0.6 x Motor parameters

Weighted sum

  • M.R.Dogar. Using Learned Affordances for Robotic Behavior Development. M.Sc. Thesis, Middle East

Technical University, Ankara, September 2007.

  • M.R.Dogar, E.Ugur, E.Sahin and M.Cakmak. Using Learned Affordances for Robotic Behavior
  • Development. . Accepted to ICRA’08.

Inspiration source: Population coding in motor cortex

slide-33
SLIDE 33

33

Using primitive behaviors Using behavioral generalization

Primitive behavior Blending of behaviors

Goal: Approach

slide-34
SLIDE 34

34

Blending Behaviors on KURT3D

Behavior blending allows to span a whole range of behaviors from a limited pre-coded primitive behaviors

slide-35
SLIDE 35

35

Part 3: Planning with Learned Affordances

slide-36
SLIDE 36

36

Behaviors

Movement Behaviors Lift Behavior

  • E.Ugur, M.R.Dogar and E.Sahin. Planning with Learned Object Affordances. Submitted to AAAI’08.
slide-37
SLIDE 37

37

Perception

slide-38
SLIDE 38

38

Predicting the next state of the entity

  • Adding the effect prototype of the behavior to be

applied to the current entity representation provides us a prediction of the expected state of the entity.

slide-39
SLIDE 39

39

Plan Generation

slide-40
SLIDE 40

40

Sample Plans

slide-41
SLIDE 41

41

Analysis of effect classes

slide-42
SLIDE 42

42

Learning Process

slide-43
SLIDE 43

43

Planning Performance

slide-44
SLIDE 44

44

Goal: Activate the Button

  • Goal:

– Any object: The range image coordinate features should be high – Button object: Mean distance should be small

slide-45
SLIDE 45

45

Planning on KURT3D

slide-46
SLIDE 46

46

Acknowledgements

Multi-sensory Autonomous Cognitive Systems

supported within the Cognitive Systems Call of FP6-IST (FP6-IST-2-004381)

Mehmet R. Dogar Emre Ugur Maya Cakmak