The Probabilistic (R)Evolution

1
The Probabilistic (R)Evolution

• Sebastian Thrun
• Director, Stanford AI Lab

2
Acknowledgements
Mike Montemerlo Dieter Fox Wolfram Burgard Nick
Roy Frank Dellaert Joelle Pineau Andreas Nuechter
team from Fraunhofer
Dave Ferguson Aaron Morris Dirk Hähnel
3
In This Talk

• I will explain the basics of probabilistic
  robotics
• I will relate probabilistic robotics to other
  robot programming paradigms
• I will discuss robot systems programmed
  probabilistically

4
Example 1 Robotics Exploration
5
(No Transcript)
6
(No Transcript)
7
(No Transcript)
8
Somerset, PA, July 02
9
(No Transcript)
10
(No Transcript)
11
Source Bureau of Abandoned Mine Reclamation, PA
12
By Scott Thayer, Red Whittaker, 16-899 MRD (CMU)
13
(No Transcript)
14
(No Transcript)
15
Groundhog
16
Courtesy of Andreas Nuecher, Joachim Hertzberg et
al
17
(No Transcript)
18
(No Transcript)
19
Example 2 Human Interaction
20
Change in Population Structure (USA)
21
Dependency Ratio per 100 working-age people (USA)
Children
Elders
22
The Nursebot Project
Providing information (TV, weather)
Monitoring compliance
Reminding (e.g., medication)
Calling help in emergencies
Facilitating inter-personal communication
Exercise assistance
Remote health services (Tele-presence)
Physical assistance
Physical Guidance
23
Nursebot Initial Prototypes (1999)
24
Field Tests
25
Escorting People
26
Characteristics

• Both systems are state-of-the-art
• Massive uncertainty
• Environment unknown, unpredictable
• People unpredictable, almost random
• Environment uncertainty drives robot behavior

27
Probabilistic Robotics
28
Probabilistic Robotics Manifesto

• p(x)

29
Example
Sensor model p(image staircase) 0.7 p(image
no staircase) 0.2
Environment prior p(staircase) 0.1

30
Probabilistic Robotics in Context

• Model-Based Robotics
• Full model, no perception
• Focus on motion planning

• Behavior-Based Robotics
• No model, no state
• Focus on environment feedback

Probabilistic Techniques

• Neuro Control
• Neural networks
• Focus on learning

• Fuzzy Control
• Fuzzy sets, fuzzy rules
• Focus on soft computing

31
Probabilistic Robotics in Context
Probabilistic Techniques
32
Robot Environment Interaction
environment

• control

• measurement

• robot belief

33
Robot Environment Interaction

• belief b

• state x

• measurement z

• belief b

• control u

• belief b

• state x

• measurement z

• belief b

. . .
. . .
34
Elements of Probabilistic Robotics
35
Markov Localization
36
Monte Carlo Localization
37
Particle Filter (1953)

• set of particles x p(x)

• belief b

• p(x)

• resample particle x with prob ? p(z x)

• p(x z)

• measurement z

• sample x p(x u,x) for each particle x

• control u

• p(x z,u)

38
Particle Filter Monte Carlo Localization
39
Monte Carlo Localization (1)
40
Monte Carlo Localization (2)
41
Particles Robustness
42
Nursebot Tracking People
43
Nursebot Tracking People

• ? robot location (particles)
• ? people location (particles)
• ? laser measurements (wall)

With Michael Montemerlo
44
Nursebot Tracking People

• ? robot location (particles)
• ? people location (particles)
• ? laser measurements (wall)

With Michael Montemerlo
45
SLAM (Simultaneous Localization and Mapping)
with particles
without particles
250 meters
46
With Dirk Hähnel, Freiburg
47
With Dirk Hähnel, Freiburg
48
The Importance of Particles
raw data
without particles
with particles
With Dirk Hähnel, Freiburg
49
250 meters
With Dirk Hähnel, Freiburg
50
Commercial Perspective Evolution Robotics
Courtesy of Paolo Pirjanian
51
Summary Thus Far

• Improved robustness for perception
• How about control ??

52
The Control Problem

• ? goal location
• ? belief (particles)
• ? true robot location

53
Robot Environment Interaction

• belief b

• p(x)

• state x

• measurement z

• belief b

• p(x z)

• control u

• f( belief )

• belief b

• p(x z,u)

• state x

• measurement z

• belief b

• p(x z,u,z)

. . .
. . .
. . .
54
Classical vs Probabilistic Control
Classical control

• control u function ( state )

Probabilistic control

• control u function ( probability
  distribution over states )

55
Probabilistic Planning

• belief p(x)

• state x

• control u

• state x2

• state x3

• state x1

• belief p2(x)

• belief p3(x)

• belief p1(x)

• measurement z1

p31(x)
p11(x)
p21(x)

• measurement z2

p32(x)
p12(x)
p22(x)

• measurement z3

p33(x)
p13(x)
p23(x)
56
Implications

• Probabilistic control good for information
  gathering
• Branching factor can be daunting

57
Greedy Exploration
58
Robot Exploration
With Reid Simmons
59
Greedy Localization
With Dieter Fox, Wolfram Burgard
60
The Fallacy of Greedy Control

• ? robot
• ? location of person
• ? belief (particles)

61
Full Probabilistic Planning Solution
With Nick Roy
62
Probabilistic Dialog System
With Joelle Pineau
63
Summary
64
Summary (1)

• Robots are inherently uncertain
• sensor limitations, environment dynamics,
• Probabilistic Robotics
• represent belief by probability distribution
• Many best known solutions are probabilistic
• Mapping, exploration, people interaction,

65
Summary (2)

• Probabilistic Robotics
• Robust
• Mathematical sound
• Easy to implement
• But Can be computationally involved

66
The Fundamental Tradeoff
Many hypotheses computationally slow robust
One hypothesis computationally fast brittle
67
What's Next?
68
Application to Autonomous Helicopters
inverted flight
Map of Gates
Prof Andrew Ng. Stanford
69
2005 DARPA Grand Challenge
70
(No Transcript)
71
www.stanfordracing.org
72
A Final
Announcement
73

• ROBOTICS SCIENCE AND SYSTEMS
• June 8-11, 2005, MIT Campus
• www.robotics-conference.org
• Single track conference, highly selective
• 9 fantastic invited plenary speakers
• Submission deadline Jan 17, 2004
• Check out www.robotics-conference.org

74
Thank You!

• More information at robots.stanford.edu