Real Robots For the Real World AAAI-1992

1
Real Robots For the Real WorldAAAI-1992

• Sebastian Thrun
• Carnegie Mellon University
• thrun_at_cs.cmu.edu

2
AI Magazine
3
(No Transcript)
4
(No Transcript)
5
Real Robots For the Real WorldAAAI-2004

• Sebastian Thrun
• Stanford AI Lab
• thrun_at_stanford.edu

6
Acknowledgements
Mike Montemerlo, Andreas Nuechter team from
Fraunhofer
7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10
Somerset, PA, July 02
11
(No Transcript)
12
(No Transcript)
13
Pennsylvania
Source Bureau of Abandoned Mine Reclamation, PA
14
Mine Mapping Systems
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
Groundhog
21
(No Transcript)
22
(No Transcript)
23
NOs

• No communication
• No human access
• No sparks
• No information on hazards
• No breadcrumbs
• No second robot

24
Mapping / SLAM
25
Simultaneous Localization And Mapping
26
Simultaneous Localization And Mapping
27
Mr. Metric
28
Extended Kalman Filter Smith/Cheeseman 86
29
space O(n2) update time O(n2) n map size
30
Real-World SLAM ResultsCourtesy of Stefan
Williams and Hugh Durrant-Whyte, Sydneysee also
Smith/Self/Cheeseman, 1986
31

• Mr. Metric
• State map robot
• Embedded in Euclidean space
• Posterior p(map, robot)
• Gaussian moments S, m
• Update time O(n2)
• Memory O(n2) with n map size
• O(1,000) features max

• Ms. Topological
• ?

32
Ms. Topological
33
Information Form Algorithms

• Offline
• Lu/Milios 97
• Bosse/Leonard 01
• Online (filter)
• Nettleton/Durrant-Whyte 02
• Thrun/Koller/Ng/Gharamani/Liu/Durrant-Whyte 03
• In-between
• Konolige/Gutmann 03

34
Inverse Correlation
35
Probability ? Information
36
From Probabilities To Information And Back(or
from metric to topological)
37
Information Matrix Interpretation
38
Theorem

• For a sparse information matrix W, the following
  updates consume constant time
• Measurement update (addition)
• Motion update (subtraction arc removal)
• Sparsification (arc removal)
• Linearization (amortized)
• But Requires approximation (See our WAFR-02
  paper)

39
(No Transcript)
40

• Mr. Metric
• State map robot
• Embedded in Euclidean space
• Posterior p(map, robot)
• Gaussian moments S, m
• Update time O(n2)
• Memory O(n2) with n map size
• O(1,000) features max

• Ms. Topological
• State maprobot
• Not embedded, all relative
• Posterior -log p(map, robot)
• Information parametrs W,x
• Update time O(1)
• Memory O(n)
• O(100,000,000) features max
• But no inference!!!

Proactive!
Lazy !
41
The Truth Out!
Metric Mapping Absolute coordinates p(x) Proacti
ve mapping
Topological Mapping Relative coordinates -log
p(x) Lazy mapping
inference
42
Data Association Problem
?
43
(No Transcript)
44
Mr. Data Association
Proactive!
Brittle !
FastSLAM
Murphy 00, Thrun et al 00, Parr et al 02,
Montemerlo et al 03, Haehnel et al 03
45
Mr. Data Association
Proactive!
Brittle !
Murphy 00, Thrun et al 00, Montemerlo et al 02,
Parr et al 02, Haehnel et al 03
46
Ms. Data Associations Insights

• Insight 1 Data association problems rare!
• Insight 2 Data association soft constraint
• Insight 3 Lazily expand data association tree

47
(No Transcript)
48
Ms. Data Association
49
Ms. Data Association
-log p 100
-log p 2
-log p 5
-log p 80
-log p 120
-log p 10
-log p 7


-log p 80
-log p 90
-log p 70
50
Data Association

• Mr. Proactive
• Particle filter many hypotheses

• Ms. Lazy
• Recursive tree search expand fringe only when
  needed

51
250 meters
Lazy recursive
Maximum likelihood
52
Before Inference
After Inference
800 meters
By Michael Montemerlo
53
Proof (Part 1 Measurement Update)
54
Proof (Part 2 Motion Update)
55
Vehicle Control
56
(No Transcript)
57
(No Transcript)
58
(No Transcript)
59
Das Experiment
60
(No Transcript)
61
(No Transcript)
62
(No Transcript)
63
Courtesy of Andreas Nuecher, Joachim Hertzberg et
al
64
(No Transcript)
65
250 meters
Bruceton Research Mine
66
(No Transcript)
67
(No Transcript)
68
(No Transcript)
69
Whats Next
70
2005 DARPA Grand Challenge
71
(No Transcript)
72
www.stanfordracing.org
73
(No Transcript)
74
Thank You!

• More information at robots.stanford.edu