Sebastian Thrun

1
Statistical Learning in RoboticsState-of-the-Art,
Challenges and Opportunities

• Sebastian Thrun
• Carnegie Mellon University

2
Main Message
Machine Learning
Robotics
3
This Talk
Robotics Research Today
Robotics Research Today
Estimation and Learning In Robotics
7 Open Problems
4
Robotics Yesterday
5
Robotics Today
6
Robotics Tomorrow?
Thanks to T. Dietterich
7
Robotics _at_ CMU, 1992
8
Robotics _at_ CMU, 1994
9
Robotics _at_ CMU 1996
With RWI / iRobot, Hans Nopper
10
Robotics _at_ CMU/UBonn, 1997
with W. Burgard, A.B. Cremers, D. Fox, D. Hähnel,
G. Lakemeyer, D. Schulz, W. Steiner
11
Robotics _at_ CMU, 1998
with M. Beetz, M. Bennewitz, W. Burgard, A.B.
Cremers, F. Dellaert, D. Fox, D. Hähnel, C.
Rosenberg, N. Roy, J. Schulte, D. Schulz
12
This Talk
Robotics Research Today
Estimation and Learning In Robotics
7 Open Problems
13
The Robot Localization Problem
?

• Position tracking (error bounded)
• Global localization (unbounded error)
• Kidnapping (recovery from failure)

14
Probabilistic Localization
Simmons/Koenig 95 Kaelbling et al 96 Burgard
et al 96 Thrun et al 96
15
Probabilistic Localization
x state t time m map z measurement u
control
Kalman 60, Rabiner 85
16
What is the Right Representation?
17
Monte Carlo Localization (MCL)
18
Monte Carlo Localization (MCL)
With Wolfram Burgard, Dieter Fox, Frank Dellaert
19
Implications for Planning Control
MDP Planner
POMDP Planner
With N. Roy
20
Monte Carlo Localization
With Frank Dellaert
21
(No Transcript)
22
(No Transcript)
23
Learning Mapsaka Simultaneous Localization and
Mapping (SLAM)
24
Learning Maps
Localization
25
Learning Maps with Extended Kalman Filters
Smith, Self, Cheeseman, 1990
26
Kalman Filter Mapping O(N2)
27
Can We Do the Same WithParticle Filters?
?
sample map pose
robot poses and maps
28
Mapping Structured Generative Model
Landmark
m1
z1
z3
measurement
. . .
s1
s2
st
s3
robot pose
u3
ut
u2
control
z2
zt
m2
With K. Murphy, B. Wegbreit and D. Koller
29
Rao-Blackwellized Particle Filters
30
Ben Wegbreits Log-Trick
31
Advantage of Structured PF Solution
Kalman O(N2)
500 features
32
3 Examples
33
Outdoor Mapping (no GPS)
With Juan Nieto, Jose Guivant, Eduardo Nebot,
Univ of Sydney
34
With Juan Nieto, Jose Guivant, Eduardo Nebot,
Univ of Sydney
35
Tracking Moving Features
With Michael Montemerlo
36
Tracking Moving Entities Through Map Differencing
37
Map-Based People Tracking
With Michael Montemerlo
38
Autonomous People Following
With Michael Montemerlo
39
Indoor Mapping

• Map point estimators (no uncertainty)
• Lazy

40
Importance of Probabilistic Component
Non-probabilistic
Probabilistic, with samples
41
Multi-Robot Exploration
DARPA TMR Maryland
DARPA TMR Texas
With Reid Simmons and Dieter Fox
42
Learning Object Models
43
Nearly Planar Maps

• Idea Exploit fact that buildings posses many
  planar surfaces
• Compacter models
• Higher Accuracy
• Good for capturing environmental change

44
Online EM and Model Selection
mostly planar map
raw data
45
Online EM and Model Selection
CMU Wean Hall
Stanford Gates Hall
46
3D Mapping Result
With Christian Martin
47
Combining Tracking and Mapping
With Dirk Hähnel, Dirk Schulz and Wolfram Burgard
48
Combining Tracking and Mapping
With Dirk Hähnel, Dirk Schulz and Wolfram Burgard
49
Underwater Mapping (with University of Sydney)

With Hugh Durrant-Whyte, Somajyoti Majunder,
Marc de Battista, Steve Scheding
50
This Talk
Robotics Research Today
Estimation and Learning In Robotics
7 Open Problems
51
Can We Learn Better Maps?
1
52
Can We Learn Control?
2
53
How Can We Learn in Context?
3

• Goal of robotics is not
• mapping
• classification
• clustering
• density estimation
• reward prediction

• But simply Doing the right thing.

54
How can we exploit Domain Knowledge in Learning?
4
55
Can we Integrating Learning and Programming?
5
56
What Can We LearnFrom Biology?
6
57
And Can We Actually DoSomething Useful?
7
58
The Nursebot Project
59
Haptic Interface (In Development)
60
Wizard of Oz Studies
By Sara Kiesler, Jenn Goetz
61
Truly Useful.?