Zürich

Roland Siegwart
Margarita Chli
Paul Furgale
Marco Hutter
Martin Rufli
Davide Scaramuzza

ETH Master Course: 151-0854-00L

Autonomous Mobile Robots
Localization I
2
Introduction
Do we need to localize or not?

 To go from A to B, does the robot need to know

where it is?

Localization I Zürich
3
Introduction
Do we need to localize or not?

 How to navigate between A and B

 navigation without hitting obstacles
 detection of goal location
 Possible by following always the left wall
 However, how to detect that the goal is reached

Localization I Zürich
4
Introduction
Do we need to localize or not?

 Following the left wall is an example of “behavior

based navigation”
 It can work in some environments but not in all
 With which accuracy and reliability do we reach the
goal?

Localization I Zürich
5
Introduction
Do we need to localize or not?

 As opposed to behavior based navigation is “map based navigation”

 Assuming that the map is known, at every time step the robot has to know where it is.
How?
 If we know the start position, we can use wheel odometry or dead reckoning. Is
this enough? What else can we use?
 But how do we represent the map for the robot?
 And how do we represent the position of the robot in the map?

Localization I Zürich
6
Introduction
Definitions
 Global localization
 The robot is not told its initial position
 Its position must be estimated from scratch

 Position Tracking
 A robot knows its initial position and “only” has to accommodate small errors in
its odometry as it moves

Localization I Zürich
7
Introduction
How to localize?

 Localization based on external sensors, beacons or landmarks

 Odometry

 Map Based Localization (without external sensors or artificial landmarks. Just use
robot onboard sensors)
 Example: Probabilistic Map Based Localization

Localization I Zürich
8 Introduction
Beacon Based Localization

Triangulation
 Ex 1: Poles with highly reflective surface and a laser for detecting them
 Ex 2: Coloured beacons and an omnidirectional camera for detecting
them (example: RoboCup or autonomous robots in tennis fields)

Localization I Zürich
9
Introduction
Beacon Based Localization

KIVA Systems, Boston (MA)

(acquired by Amazon in 2011)

Unique marker with known absolute

2D position in the map

Prof. Raff D'Andrea, ETH

Localization I Zürich
10
Introduction
Motion Capture Systems

 High resolution (from VGA up to 16 Mpixels)

 Very high frame rate (several hundreds of Hz)
 Good for ground truth reference and multi-robot control strategies
 Popular brands:
 VICON (10kCHF per camera),
 OptiTrack (2kCHF per camera)

Localization I Zürich
11
Introduction
Map-based localization
 Consider a mobile robot moving in a known environment.

Localization I Zürich
12
Introduction
Map-based localization
 Consider a mobile robot moving in a known environment.
 As it starts to move, say from a precisely known location, it can keep track of its motion
using odometry.

Localization I Zürich
13
Introduction
Map-based localization
 Consider a mobile robot moving in a known environment.
 As it starts to move, say from a precisely known location, it can keep track of its motion
using odometry.

Localization I Zürich
14
Introduction
Map-based localization
 Consider a mobile robot moving in a known environment.
 As it starts to move, say from a precisely known location, it can keep track of its motion
using odometry.

Sensor reference frame

Localization I Zürich
15
Introduction
Map-based localization
 Consider a mobile robot moving in a known environment.
 As it starts to move, say from a precisely known location, it can keep track of its motion
using odometry.
 The robot makes an observation and updates its position and uncertainty

Sensor reference frame

Localization I Zürich
16
Ingredients
Probabilistic Map-based localization

 Probability theory → error propagation, sensor fusion

 Belief representation (map/position) → discrete / continuous

 Motion model → odometry model

 Sensing → measurement model

Localization I Zürich
17
Probabilistic localization
belief representation
‫݌‬ሺ‫ݔ‬ሻ

a) Continuous map with Kalman Filter

Localization
single hypothesis probability
distribution ‫݌‬ሺ‫ݔ‬ሻ ‫ݔ‬
‫݌‬ሺ‫ݔ‬ሻ

b) Continuous map with

multiple hypotheses
probability distribution ‫݌‬ሺ‫ݔ‬ሻ ‫ݔ‬
‫݌‬ሺ‫ݔ‬ሻ

c) Discretized metric map (grid Markov Localization

݇) with
probability distribution ‫݌‬ሺ݇ሻ ݇
‫݌‬ሺ‫ݔ‬ሻ

d) Discretized topological map

(nodes ݊) with
probability distribution ‫݌‬ሺ݊ሻ ݊
A B C D E F G

Localization I Zürich
18
Belief Representation
Characteristics

 Continuous  Discrete

 Precision bound by sensor data  Precision bound by resolution of

 Typically single hypothesis pose discretisation
estimate  Typically multiple hypothesis pose
 Lost when diverging (for single estimate
hypothesis)  Never lost (when diverges
 Compact representation and converges to another cell)
typically reasonable in processing  Important memory and processing
power. power needed. (not the case for
topological maps)

Localization I Zürich
19
Odometry

 Definition
 Dead reckoning (also deduced reckoning or odometry) is the process of
calculating vehicle's current position by using a previously determined position
and estimated speeds over the elapsed time

 Robot motion is recovered by integrating proprioceptive sensor velocities

readings
 Pros: Straightforward
 Cons: Errors are integrated -> unbound

 Heading sensors (e.g., gyroscope) help to reduce the accumulated errors

but drift remains

Localization I Zürich
20
Odometry
The Differential Drive Robot (1)

 x  x 
x   y xˆt  xt 1   y   f ( xt 1, ut )
   

x
Localization I Zürich
21
Odometry
Wheel Odometry
 Kinematics

  
 s cos(  ) This term comes from the application
 xt 1   2 of the Instantaneous Center of Rotation
   
xˆt  f ( xt 1, ut )   yt 1    s sin(  ) 
 2 
t 1    
  Can you demonstrate these equations?
s  sl
s  r
2

sr  sl
 
b
Localization I Zürich
22
Odometry
Odometric Error Propagation
 Error model

Pt  Fxt1   xt1  Fxt1  FS  S  FS

T T

kr sr 0 
S   
 0 kl sl 

Fxt1  f xt1

FS

Localization I Zürich
23 Odometry
Growth of Pose uncertainty for Straight Line
Movement
 Note: Errors perpendicular to the direction of movement are growing much faster!

Localization I Zürich
24 Odometry
Growth of Pose uncertainty for Movement on a
Circle
 Note: Errors ellipse does not remain perpendicular to the direction of movement!

Localization I Zürich
25
Odometry
Example of non-Gaussian error model

 Note: Errors are not shaped like ellipses!

Courtesy AI Lab, Stanford

[Fox, Thrun, Burgard, Dellaert, 2000]

Localization I Zürich
26
Odometry
Error sources

Deterministic Non-Deterministic
(Systematic) (Non-Systematic)

 Deterministic errors can be eliminated by proper calibration of the system.

 Non-Deterministic errors are random errors. They have to be described by error
models and will always lead to uncertain position estimate.

 Major Error Sources in Odometry:

 Limited resolution during integration (time increments, measurement resolution)
 Misalignment of the wheels (deterministic)
 Unequal wheel diameter (deterministic)
 Variation in the contact point of the wheel (non deterministic)
 Unequal floor contact (slippage, non planar …) (non deterministic)

Localization I Zürich
27
Odometry
Calibration of systematic errors [Borenstein 1996]
 The unidirectional square path experiment

 BILD 1 Borenstein

Localization I Zürich
28
Odometry
Calibration of Errors II (Borenstein [5])
 The bi-directional square path experiment

 BILD 2/3 Borenstein

Localization I Zürich