An Introductory Discussion on Robot Navigation

1
An Introductory Discussion on Robot Navigation

• by Nicola Bellotto
• Centre for Hybrid Intelligent Systems
• University of Sunderland

2
Introduction

• Navigation in Mobile Robotics
• The Localization task
• Place recognition using vision
• Odometry integration
• Experimental results
• Future guidelines

3
Navigation in Mobile Robotics
4
The Navigation tasks
The map can be provided by the user. If not, the
learning can be supervised or completely
autonomous.
(Map-learning)
Localization
The robot must know its position within the
environment.
The goal should be reached as soon as possible
and at the minimum cost.
Path-planning
5
Map

• Metric positions and objects are placed using
  (x, y) coordinates in a common 2D frame

• Topological graph where vertices are
  positions/objects and edges are spatial relations
  among them

window
door
Y
y
desk
x
printer
X
6
Localization

• Three kinds of localization strategies
• direct position inference
• single-hypotesis tracking
• multiple-hypotesis tracking
• For each strategy, three different ways to
  represent maps and position
• metric map / metric position
• topological map / metric position
• topological map / topological position

7
Localization
Direct position inference
Single-hypotesis tracking
Multiple-hypotesis tracking
8
Localization
Extracted from Filliat, D., Meyer, J. A.
(2003). Map-based navigation in mobile robots I.
A review of localization strategies. Cognitive
Systems Research 4, pp. 243-282
9
Path-planning

• Given the current position of the robot and the
  target destination (goal), find on the map the
  best path connecting the two locations.
• Usually, the problem is solved discretizing the
  space and applying a standard graph algorithm
  (e.g. Dijkstra, A)
• Must be updated in real-time to handle changes in
  the real environment (e.g. there is an unexpected
  obstacle or the goal has moved)

10
The Localization task
11
Localization strategy

• Direct position inference does not require
  idiothetic data, but suffers of perceptual
  aliasing.
• Single-hypotesis requires both idiothetic and
  allothetic data, keeping track of the last
  estimated position, but fails in case this last
  one is completely wrong.
• Multiple-hypotesis like the previous one, but
  keeps track of all the possible positions,
  updating them in parallel.

12
Map / Position representation

• Metric map / metric position very precise but
  computational expensive. Needs a reliable sensor
  model.
• Topological map / metric position does not need
  a sensor model, but still computational expensive
  because of the continuos position estimation.
• Topological map / topological position like the
  previous one, but with a less expensive discrete
  position estimation.

13
Comparison
Extracted from Filliat, D., Meyer, J. A.
(2003). Map-based navigation in mobile robots I.
A review of localization strategies. Cognitive
Systems Research 4, pp. 243-282
14
The robot-waiter scenario

• Perceptual aliasing (e.g. two identical doors in
  different places)
• Robot completely lost (e.g. surrounded by the
  folk)

• Not necessary to know the exact position (in
  terms of centimeters)
• Path-planning fast enough to handle a highly
  dynamic environment

Multiple-hypotesis localization
Topological map / topological position
15
Place recognition using vision
16
Techniques

• Artificial landmarks
• Landmarks placed by humans
• Need a modification of the environment
• Often impossible to arrange
• Natural landmarks
• Landmarks already present in the environment
• The recognition is challenging
• It is not just a single object (e.g. a door),
  could be a wide group of different objects (e.g.
  a room)

17
A place recognition algorithm

• Using a normal camera, reconstruct a panoramic
  view (360) for every single place.
• Given a new snapshot from one place, identify at
  which panoramic image (which place) it belongs
  to.
• The point above include also the offset of the
  current snapshot with regard to a reference angle
  0 of the panoramic image

18
The algorithm - 1st step
Given a new snapshot from the camera and a
panoramic view of the place divide the snapshot
in vertical slots for each slot for each
pixel-column of the panoramic image save the
number N of the slot that fits better and the
relative matching value M endfor endfor
19
The algorithm - 1st step
Current snapshot
N 2 2 3 2 4 4 1 4 3 3 3 1 4 2 2 1 3 4 1 1 2 2 3
3 4 4 2 2 2 4 1 1 2 2 3 1 2 3 4 4 4 1 4 3
Panoramic image
M m0 m1 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . mW
20
The algorithm - 2nd step
Given the two vectors N and M calculated
before for each element x of N for each element
i in x-SLOT_W, xSLOT_W for each slot
number if slot number equal x sum the
relative value of M endif endfor the sum
is weighted depending on the distance from
x endfor store the best sum endfor
21
The algorithm - 2nd step
if

then
x
best match and relative best index
weight
22
An example

• Given a sequence of snapshots taken every 15...

... the resulting panoramic view is like this
23
Odometry integration
24
Dead reackoning

• Relative position estimation based on internal
  measures of the movement, normally using
  encoders.
• Not reliable for long distances because subject
  to cumulative error.
• Necessary external sensors to observe useful
  features from the environment to correct the
  error.

25
Cumulative error
Real path
Measured path
26
On the topological map

• The idea is to reset the odometry information
  each time a new place is identified.
• Several methods are adopdet, mainly based on
  probability distributions.

27
Experimental results
28
The scenario

• Neuro Robots Laboratory
• Four corners of a square ( 1m x 1m)

Corner 4
Corner 3
Corner 2
Corner 1
29
Set-up

• For each corner, one snapshot every 15.
• Reconstruction of four panoramic images using the
  previous algorithm (matching function adopted
  Normalized Correlation Coefficient)
• Every panoramic image saved into the relative
  node of the graph.

30
Experiments

• Use of panoramic images obtained a few days
  before.
• Small changes in the environment (chairs in
  different positions, different objects on the
  desks, ...).
• MIRA moved randomly among the four corners,
  taking a new snapshot at interval of 1s.

31
Results

• MIRA could correctly identify the current corner
  in almost every case and every orientation.
• In case of wrong estimation, the associated
  probability was very low and, anyway, the correct
  position could be restored just turning the robot
  a few degrees.

32
Results

• The algorithm of place recognition seemed to be
  quite robust to moving obstacles (e.g. a person
  walking or a door opening).

33
Comments

• The biggest constraint is the illuminance of the
  environment. Same experiments conducted with
  different light conditions reported bad results.
• The problem could be resolved with an appropriate
  filtering of the image (e.g. histogram
  equalization, edge detector).

34
Comments

• Beyond a distance of 20-30cm from the corner,
  MIRA could not locate itself correctly.
• This problem can be resolved using the odometry
  information.

35
Future guidelines
36
MUST be done

• Integration of the odometry information.
• Improvement of the place recognition algorithm
  thus to be robust enough to different light
  conditions
• Some kind of map-learning to skip the first part
  of image acquisition

37
MIGHT be done

• Different kind of place recognition (features
  extraction, neural networks, ...)
• Omnidirectional vision sensor, alone or combined
  with the normal camera

38
Some references

• Papers
• Filliat, D., Meyer, J.A. (2003). Map-based
  navigation in mobile robots I. A review of
  localization strategies. Cognitive Systems
  Research 4, pp. 243282.
• Meyer, J.A., Filliat, D. (2003). Map-based
  navigation in mobile robots II. A review of
  map-learning and path-planning strategies.
  Cognitive Systems Research 4, pp. 283317.
• Books
• Murphy, R. R. (2000). Introduction to AI
  Robotics. The MIT Press
• Siek, J., Lee, L-Q., Lumsdaine, A. (2002). The
  Boost Graph Library. Addison-Wesley
• Web
• C Boost Libraries - http//www.boost.org
• Intel Open Source Computer Vision Library -
  http//sourceforge.net/projects/opencvlibrary

39
Questions Answers