BehaviorBased Landmarks for Topological Mapping and Navigation of Indoor Environments

1
Behavior-Based Landmarks for Topological Mapping
and Navigation of Indoor Environments

• by
• Curtis Nielsen

2
Outline

• Introduction
• Behavior-based landmarks
• Localization
• Map-building
• Conclusion

3
Introduction

• Big picture Search and rescue
• Important step Mapping and localization
• Previous Work
• Thrun
• Geometric maps are used to produce topological
  maps
• or human input is used to denote important places
• Choset and Nagatani
• Topological maps Generalized Voronoi Graph
• Dedeoglu, Mataric, and Sukhatme
• Topological maps Sonar, drift-stabilized gyro,
  compass, pan-tilt color camera
• Gibson and Norman
• Affordances What can be done in the environment?

4
Our Goals

• Inexpensive robots
• Minimal computational requirements
• Intuitive topological representation

5
Behavior-based Landmarks

• Set of behaviors afforded to the robot based on
  what the environment allows the robot to do
  (Gibsonian affordances)
• Behaviors are loosely based on navigational
  primitives such as turn right, turn left, go
  forward.
• Behaviors are found using sonar measurements.

6
Behavior-Based LandmarksAfforded Behaviors
Temporal Average
Spatial Average
Local Maxima
Vector Summation
7
Localization

• The ability of the robot to learn where it is in
  a known environment
• Extension of Thruns work to topological domains.
• Topology based on behaviors.
• Required for simultaneous localization and
  mapping.

8
Localization Algorithm

• Probability of the location of the robot at time
  t.
• Calculation of a-values The probability of the
  location of the robot given what it did and saw
  before this landmark.
• Calculation of ß-values The probability of the
  location of the robot given what it did and saw
  after this landmark

9
Localization t 1
ß
a
ID
Time step
10
Localization t 2
ß
a
ID
Time step
11
Localization t 3
ß
a
ID
Time step
12
Localization t n
ß
a
ID
Time step
13
Other localization experiments
Real-world (3rd floor of the Talmage building)
14
Landmark Classification Error

• 5 misclassification
• 10 misclassification
• 20 misclassification

15
Building a behavior-based map

• Landmark disambiguation
• Sealing algorithm
• Exploration
• Closing an area
• Verification
• Mapping algorithm

16
Landmark Disambiguation

• The Problem Distinguish between similar
  appearing places in the environment
• Typically a hard problem for topological
  map-building algorithms due to error in robot
  movement

• A, B, and C are all T-shaped landmarks and have a
  similar topological appearance but only A and C
  represent the same location in the environment

17
Sealing Algorithm Exploration

• First Find the nearest landmark with an
  unexplored behavior (Dijkstras algorithm)

• Second Use a wall-following heuristic to
  estimate which unexplored behavior will most
  likely lead to a previously explored landmark

18
Exploration heuristic in action
19
Sealing Algorithm Closing an area

• Recognize when the robot has reached a previously
  discovered landmark.

• SolutionA bounding box surrounding the possible
  positions of where the landmark could be.

20
Finding a local landmark that could match.
21
Sealing Algorithm Verification

• Can not tell if landmark A landmark B
• Heuristic
• If As neighbors match Bs neighbors
• Then the landmarks match.
• Choose neighbors using shortest-path planning

22
Verification Algorithm in action
23
Mapping Algorithm

• Move to a landmark and classify it
• Determine the position of the new landmark
• Determine if the new landmark could match a
  previous landmark
• Choose a behavior to follow
• Continue until the map is finished

24
3rd Floor of the Talmage Building
25
3rd Floor of the Talmage Building
26
3rd Floor of the Talmage Building
27
Other Map-building Experiments
28
When to use which algorithm?
29
Conclusions

• Summary
• We have designed a system that uses inexpensive
  robotics and minimal computational requirements
  to represent orthogonal environments with a
  topological map.
• We have validated our localization and
  map-building with experiments in the real-world
  and in simulation.

• Future Work
• Extend our work to non-orthogonal environments.
• Human-robot interaction
• Sketch-based interfaces

30
Questions?