Probabilistic Techniques for Mobile Robot Navigation

1
Probabilistic Techniquesfor Mobile Robot
Navigation
Wolfram Burgard University of Freiburg
Department of Computer Science http//www.informa
tik.uni-freiburg.de/burgard/
2
Robotics Today
3
DARPA Grand Challenge
Courtesy by Sebastian Thrun
4
Robot Projects Interactive Tour-guides
Rhino
Albert
Minerva
5
Probabilistic Techniques in Robotics

• Perception state estimation
• Action utility maximization
• Key Question
• How to scale to higher-dimensional spaces

6
Nature of Data
Odometry Data
7
Platforms
8
Robot Projects Acting in the Three-dimensional
World
Herbert
Zora
Groundhog
9
Dimensions of Mobile Robot Navigation
SLAM
localization
mapping
integrated approaches
active localization
exploration
motion control
10
Dimensions of Mobile Robot Navigation
SLAM
localization
mapping
integrated approaches
active localization
exploration
motion control
11
Outline

• Localization
• Mapping
• Exploration

12
Probabilistic Localization
13
Mobile Robot Localization with Particle Filters
14
MCL Sensor Update
15
PF Robot Motion

16
Particle Filter Algorithm
4. Re-sample
17
MCL Global Localization (Sonar)
18
Vision-based Localization
19
Dimensions of Mobile Robot Navigation
SLAM
localization
mapping
integrated approaches
active localization
exploration
motion control
20
Outline

• Localization
• Mapping
• Exploration

21
Simultaneous Localization and Mapping (SLAM)

• To determine its position, the robot needs a map.
• During mapping, the robot needs to know its
  position to learn a consistent model
• Simultaneous localization and mapping (SLAM) is a
  chicken and egg problem

22
Why SLAM is Hard Raw Odometry
23
Scan Matching

• Maximize the likelihood of the i-th pose and map
  relative to the (i-1)-th pose and map.

24
Mapping using Scan Matching
25
Probabilistic Formulation of SLAM
26
Key Question/Problem

• How to maintain multiple map and pose hypotheses
  during mapping?
• Ambiguity caused by the data association problem.

27
A Graphical Model for SLAM
28
Rao-Blackwellized Particle Filters for SLAM

• Observation
• Given the true trajectory of the robot, we can
  efficiently compute the map (mapping with known
  poses).
• Idea
• Use a particle filter to represent potential
  trajectories of the robot.
• Each particle carries its own map.
• Each particle survives with a probability that is
  proportional to the likelihood of the observation
  given that particle and its map.

Murphy et al., 99
29
Factorization Underlying Rao-Blackwellization
Mapping with known poses
Particle filter representing trajectory hypotheses
30
Example
3 particles
31
Limitations

• A huge number of particles is required.
• This introduces enormous memory and computational
  requirements.
• It prevents the approach from being applicable in
  realistic scenarios.

32
Challenge

• Reduction of the number of particles.
• Approaches
• Focused proposal distributions (keep the samples
  in the right place)
• Adaptive re-sampling (avoid depletion of
  relevant particles)

33
Motion Model for Scan Matching
Raw Odometry
Scan Matching
34
Incorporating the Current Measurement
End of a corridor Free space Corridor
35
Application Example
Loop Closure
36
Application Example
Loop Closure
37
Application Example
38
Map of the Intel Lab

• 15 particles
• four times faster than real-timeP4, 2.8GHz
• 5cm resolution during scan matching
• 1cm resolution in final map

39
Outdoor Campus Map

• 30 particles
• 250x250m2
• 1.75 km (odometry)
• 20cm resolution during scan matching
• 30cm resolution in final map

40
Multi-Level Surface Maps

• Map size 195 by 146 m
• Cell resolution 10 cm
• Number of data points 20,207,000

41
Resulting Map
42
MIT Kilian Court
43
MIT Kilian Court
44
Dimensions of Mobile Robot Navigation
SLAM
localization
mapping
integrated approaches
active localization
exploration
motion control
45
Exploration

• The approaches seen so far are purely passive.
• Given an unknown environment, how can we control
  the robot(s) to efficiently learn a map?
• By reasoning about control, the mapping process
  can be made much more effective.

46
Decision-Theoretic Formulation of Exploration
cost (path length)
reward (expected information gain)
47
Exploration with Known Poses
48
Multi-Robot Exploration

• How to control teams of robots to effectively
  cover an unknown environment?
• How to deal with the complexity of the assignment
  problem?
• How to deal with limitedcommunication ranges?

49
Levels of Coordination

• No exchange of information
• Implicit coordination Sharing a joint map
• Communication of the individual maps and poses
• Central mapping system
• Explicit coordination Determine better target
  locations to distribute the robots
• Central planner for target point assignment

50
Example
Second robot
First robot
51
Idea

• Choose target locations at the frontier to the
  unexplored area by trading off the expected
  information gain and travel costs.
• Reduce utility of target locations whenever they
  are expected to be covered by another robot.

52
Coordinated Multi-Robot Exploration

• Determine the frontier cells.
• Compute for each robot the cost for reaching each
  frontier cell.
• Choose the robot with the optimal overall
  evaluation and assign the corresponding target
  point to it.
• Reduce the utility of the frontier cells visible
  from that target point.
• If there is one robot left go to 3.

53
Application Example
First robot
Second robot
54
Real World Experiment
55
Real World Experiment
56
A Simulated Run
Independent robots
Coordinated robots
57
Extension to Systems with Limited Communication
Maps are only communicated within clusters
58
Combining SLAM and Exploration
59
Where to Move Next?
60
Naïve Approach to Combine Exploration and Mapping

• Learn the map using a Rao-Blackwellized particle
  filter.
• Apply an exploration approach that minimizes the
  map uncertainty.

61
Disadvantage of the Naïve Approach

• Exploration techniques only consider the map
  uncertainty for generating controls.
• They avoid re-visiting known areas.
• Data association becomes harder.
• More particles are needed to learn a correct map.

62
Application Example
Path estimated by the particle filter
True map and trajectory
63
Map and Pose Uncertainty
pose uncertainty
map uncertainty
64
Goal

• Integrated approach that considers
• exploratory actions,
• place revisiting actions, and
• loop closing actions
• to control the robot.

65
Dual Representation for Loop Detection

• Trajectory graph stores the path traversed by the
  robot.
• Occupancy grid map represents the space covered
  by the sensors.
• Loops correspond to long paths in the trajectory
  graph and short paths in the geometric map.

66
Example Trajectory Graph
67
Application Example
68
Real Exploration Example
69
Comparison
Map uncertainty only
Map and pose uncertainty
70
Quantitative Results
Localization error
avg. localization error m
map uncertainty
map and pose uncertainty
71
Corridor Exploration
72
Example Entropy Evolution
73
Summary

• Probabilistic techniques are a powerful tool for
  solving fundamental robot navigation problems
• One key challenge lies in the high dimensionality
  of the underlying state estimation problems
• They can be addressed by
• developing appropriate representations and
  algorithms and by
• actively controlling the robots.

74
Future Work

• Learning
• Utilizing background knowledge in state
  estimation problems
• Applications
• Integrated/embedded systems
• Human-robot interaction

75
Place Labeling

• How to classify places?

76
Online Categorization of Places
77
Mapping in Dynamic Environments
78
Tracking People
79
Advanced Security for (Autonomous Cars)