+ All Categories
Home > Documents > From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand...

From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand...

Date post: 20-Jan-2016
Category:
Upload: leo-robertson
View: 214 times
Download: 0 times
Share this document with a friend
52
From Cognitive From Cognitive Spatial Spatial Mapping to Mapping to Robot Mapping Robot Mapping Margaret Jefferies Margaret Jefferies University of Waikato University of Waikato New Zealand New Zealand Hans-Wissenschaftskolleg Hans-Wissenschaftskolleg University of Bremen University of Bremen Germany Germany
Transcript
Page 1: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

From Cognitive From Cognitive Spatial Mapping Spatial Mapping

to Robot Mappingto Robot Mapping

Margaret JefferiesMargaret JefferiesUniversity of WaikatoUniversity of Waikato

New ZealandNew ZealandHans-WissenschaftskollegHans-Wissenschaftskolleg

University of BremenUniversity of BremenGermanyGermany

Page 2: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Autonomous Mobile RobotsAutonomous Mobile Robots

Page 3: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 4: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 5: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 6: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Robots might not be taking over the world any time soon but they could soon rule the roost if most New Zealanders have their way.

More than two-thirds of New Zealanders would welcome robots to do chores around the house, according to a study of 750 people, commissioned by Honda. Most people wanted robots to help with housework, many wanted an extra mechanical hand with the washing up and some wanted a robot to mow the lawns.

Page 7: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Politicians and the All Blacks need to watch their backs – some respondents suggested robots should replace politicians and that a team of robots might fare better than the present rugby team.

Some people said they would even swap their partners for robots. Women were keener for a robotic partner, with 5.5 per cent saying they would like to switch, compared with just 3.3 per cent of men wanting to replace their partner.

Page 8: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Autonomous Mobile RobotsAutonomous Mobile RobotsMappingMapping

Robot computes its own map from it own Robot computes its own map from it own

experience of its environment with its experience of its environment with its

imperfect sensors and imperfect odometryimperfect sensors and imperfect odometry

What’s the problem?What’s the problem?

Page 9: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

DemoDemo

Page 10: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Simultaneous Localisation and Simultaneous Localisation and Mapping (SLAM)Mapping (SLAM)

Robot needs to estimate its location at the same Robot needs to estimate its location at the same

time it is estimating its maptime it is estimating its map

Page 11: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

The localisation problem

Page 12: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 13: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 14: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

ApproachesApproaches

• Absolute Metric Mapping Absolute Metric Mapping

(Global metric mapping)(Global metric mapping)

• Topological MappingTopological Mapping

(Local metric maps)(Local metric maps)

Page 15: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Representation Representation Global MapsGlobal Maps

• Global evidence-grid approach Global evidence-grid approach

Page 16: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 17: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Global metric map

Page 18: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Global metric map

Page 19: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

The Correspondence Problem (Closing the Cycle)

Page 20: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Topological RepresentationsTopological Representations

Page 21: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Topological RepresentationsTopological Representations

Page 22: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

The Correspondence Problem (Closing the Cycle)

Page 23: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

From Cognitive Spatial Mapping From Cognitive Spatial Mapping to Robot Mappingto Robot Mapping

Cognitive MapCognitive Map

An agents (human animal or robot’s) An agents (human animal or robot’s) memory of the spatial environmentmemory of the spatial environment

Page 24: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

From Cognitive Spatial Mapping From Cognitive Spatial Mapping to Robot Mappingto Robot Mapping

• Draw inspiration from the way in which Draw inspiration from the way in which

humans and animals solve similar problemshumans and animals solve similar problems

• Study the way humans and animals solve Study the way humans and animals solve

similar spatial mapping problems (to robots)similar spatial mapping problems (to robots)

Page 25: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

The Local SpaceThe Local Space

• The space that appears to enclose the viewerThe space that appears to enclose the viewer

• Initial notion of “where am I” Initial notion of “where am I”

• A container where objects are located and A container where objects are located and

where actions take placewhere actions take place

Page 26: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Bounded SpaceBounded Space

O’Keefe and Burgess Nature 1996 - hippocampus

Page 27: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Bounded spaceBounded space

• Russell Epstein and Nancy KanwisherRussell Epstein and Nancy Kanwisher– Nature (1998), Neuron (1999)Nature (1998), Neuron (1999)

• Parahippocampus encodes the layout of the Parahippocampus encodes the layout of the local space – the enclosed spacelocal space – the enclosed space

Page 28: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Bounded spaceBounded space

• Environmental Psychologists / GeographersEnvironmental Psychologists / Geographers

• 1980’s work of the Kaplans1980’s work of the Kaplans

• Stamps and Smith (2004)Stamps and Smith (2004)

Page 29: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

The Local Space is The Local Space is GeometricGeometric

• Ken Cheng Ken Cheng – Cognition (1986)Cognition (1986)

• Margules and Gallistel Margules and Gallistel – Animal Learning and Behavior(1988)Animal Learning and Behavior(1988)

• Huttenlocher et alHuttenlocher et al– Cognitive Psychology (1979), (1994) Cognitive Psychology (1979), (1994)

• Hermer and Spelke Hermer and Spelke – Nature (1994), Cognition (1996)Nature (1994), Cognition (1996)

Page 30: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Exits are importantExits are important

• Evolutionary psychologistsEvolutionary psychologists– Kaplans (1980s)Kaplans (1980s)– Laslo et al “the Evolution of Cognitive Maps” Laslo et al “the Evolution of Cognitive Maps”

(1993)(1993)

• Environmental psychologistsEnvironmental psychologists– Herzog (2001 – 2004)Herzog (2001 – 2004)– Visual accessVisual access

Page 31: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

The Local Mapping ApproachThe Local Mapping Approach

Page 32: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

E1 E2

unknown

The Local Mapping ApproachThe Local Mapping Approach

Occlusion Map Local space representation

Page 33: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

E1E2

E3

E4

Page 34: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Putting it all togetherPutting it all together

• The Theory of Siegel and White has The Theory of Siegel and White has

dominated thinking in this area since dominated thinking in this area since

it was first proposed in 1975it was first proposed in 1975

landmarkroute / topological survey

global metric

Most computational cognitive mapping approaches use all of these

Page 35: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Putting it all togetherPutting it all together

• The Theory of Siegel and White has The Theory of Siegel and White has

dominated thinking in this area since dominated thinking in this area since

it was first proposed in 1975it was first proposed in 1975

landmarkroute / topological survey

global metric

Most computational cognitive mapping approaches use all of these

Page 36: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

E1

E2E3

E4

1

E1E2

E3

2

E1E2

E3

2

1

2

E2

E3

1

2

E2

E3

Local space

Topological Map Global Metric Map

Page 37: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 38: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

DemoDemo

Page 39: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Detecting Cycles in a Global Metric MapDetecting Cycles in a Global Metric Map

Page 40: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 41: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

• Need to figure out if a newly encountered Need to figure out if a newly encountered

local space is already in the topological local space is already in the topological

map map

• Need to account for the uncertainty in local Need to account for the uncertainty in local

spacespace– In particular occlusionIn particular occlusion

• Want to do it quicklyWant to do it quickly

Closing Cycles in a Topological Closing Cycles in a Topological MapMap

2D Landmarks2D Landmarks

Page 42: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Closing Cycles in a Topological Closing Cycles in a Topological MapMap

2D Landmarks2D Landmarks

• Find (eventually) a signature that identifies Find (eventually) a signature that identifies

the the local space from wherever it is the the local space from wherever it is

approachedapproached

• Learn what it is that makes each local Learn what it is that makes each local

space different from all the othersspace different from all the others

• Whenever we compute a new local space Whenever we compute a new local space

we match it against these signatureswe match it against these signatures

Page 43: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Signature learningSignature learning• A backprop Neural Network

• Feature selection

• Input values are discretised into intervals (200mm) and 45o

• Classification – Output values between 0 and 1 indicate the degree of similarity

12

3

4 5

6

71

2

3

4 5

6

7

Page 44: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

11

2

34

5

6

7

8

9

10

11

Matches 2

Page 45: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Local Local

spacespace11 22 33 44 55 66 77 88 99 1010 1111

PredictionPrediction .78.78 .94.94 .89.89 .71.71 -.11-.11 .72.72 .18.18 .51.51 .34.34 .36.36 .04.04

2*1

2

34

5

6

7

8

9

1011

Page 46: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

DisadvantageDisadvantage

• NN doesn’t tell us how the local spaces match just that NN doesn’t tell us how the local spaces match just that

they do.they do.

• Need to find the connectivityNeed to find the connectivity

Page 47: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

ASRASR 11 22 33 44 55 66 77 88 99 1010

PredictionPrediction .46.46 .97.97 .91.91 .48.48 .64.64 .26.26 .57.57 .88.88 .15.15 .77.77

1

2

3

45

6

7

89

10

Best prediction is for 2

Should be 3

Page 48: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

3D Visual Landmarks3D Visual Landmarks

Page 49: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 50: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.
Page 51: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

ConclusionConclusion

• Recognising places they have been to Recognising places they have been to before is a hard problem for robotsbefore is a hard problem for robots

• There is no perfect solution!There is no perfect solution!

• Then there is the dynamics of the Then there is the dynamics of the environment to contend withenvironment to contend with

Page 52: From Cognitive Spatial Mapping to Robot Mapping Margaret Jefferies University of Waikato New Zealand Hans-Wissenschaftskolleg University of Bremen Germany.

Recommended