Mapping and Localization for Robots

The Occupancy Grid Approach

Agenda

Introduction Mapping

Occupancy grids Sonar Sensor Model Dynamically Expanding Occupancy Grids

Localization Iconic Feature-based Monte Carlo

An intelligent robot is a mechanical creature which can function autonomously.

Intelligent – the robot does not do things in a mindless, repetitive way.

Function autonomously – the robot can operate in a self-contained manner, under reasonable conditions, without interference by a human operator.

Robots in museums

Personal Robots

Robots in space

The problem of Navigation

Where am I going? What’s the best way there? Where have I been? Where am I? How am I going to get there?

Mapping

Topological Mapping Features and Landmarks Milestones with connections Hard to scale

Metric Mapping Geometric representations Occupancy Grids Larger maps much more computationally intensive

Map Making

Demo of Mapping

The Littlejohn Project http://www-2.cs.cmu.edu/afs/cs.cmu.edu/project/mercator/www

/

littlejohn/

Occupancy Grids

A tool to construct an internal model of static environments based on sensor data.

The environment to be mapped is divided into regions.

Each grid cell is an element and represents an area of the environment.

Representation of Occupancy Grids

Sonar Sensor Model

Methods of Sonar Reading

Probabilistic Methods: Bayesian Dempster-Shafer HIMM (Histogrammic In Motion Mapping)

Why Probabilistic Mapping?

Noise in commands and sensors Commands are not executed exactly

(eg. Slippage leads to odometry errors) Sonars have several error issues

(eg. cross-talk, foreshortening, specular reflection)

Occupancy Grids

Pros Simple Accurate

Cons Require fixed-size environment:

difficult to update if size of mapped area changes.

Dynamically Expanding Occupancy Grids

Variable-sized maps Ability to increase size of map, if new areas

are added to the environment Start mapping at center of nine-block grid As robot explores, new cells are added Global map is stored outside the RAM in a

file or a database

Representation of DEOGs

Adding Cells to a DEOG

Dynamically Expanding Occupancy Grids

Best (the only?) solution for mapping changing environments.

Saves RAM Other useful information can be stored in the

map More complicated to program than regular

occupancy grids

Localization

Where am I?

Methods: Iconic Feature-based Monte-Carlo

Iconic Localization

Use raw sensor data Uses occupancy grids Current map is compared with original map.

If original map has errors, localization is very inaccurate.

Localization errors accumulate over time.

The Concept

“pose”: (x, y, θ)location, orientation

Compare small local occupancy grid with stored global occupancy grid.

Best fit pose is correct pose.

Feature-based Localization

Compares currently extracted features with features marked in a map.

Requires presence of easily extractable features in the environment.

If features are not easily distinguishable, may mistake one for the other.

Monte Carlo Localization

Probabilistic 1. Start with a uniform distribution of possible poses (x, y,

) 2. Compute the probability of each pose given current

sensor data and a map 3. Normalize probabilities

Throw out low probability points Performance

Excellent in mapped environments Need non-symmetric geometries

References:

Introduction to AI RoboticsDr. Robin Murphy

Dynamically Expanding Occupancy GridsBharani K. Ellore

Multi-agent mapping using dynamic allocation utilizing a storage systemLaura Barnes, Richard Garcia, Todd Quasny, Dr. Larry Pyeatt

Robotic Mapping: A surveySebastian Thrun

Littlejohn Projecthttp://www-2.cs.cmu.edu/afs/cs.cmu.edu/project/mercator/www/littlejohn/

CYE www.prorobotics.com The Honda Asimo http://asimo.honda.com Mars Rover http://marsrovers.jpl.nasa.gov/home/