Building Accurate Maps using Rao-Blackwellized Particle Filters

89 Pages·2009·5.58 MB·English

Building Accurate Maps
using Rao-Blackwellized
Particle Filters
Cyrill Stachniss
University of Freiburg, Germany
Thanks and partial slide courtesy of Mike Montemerlo and Dirk Haehnel
Research Lab for Autonomous
Intelligent Systems
(cid:1) Headed by Prof. Dr. Wolfram Burgard
(cid:1) 1 academic advisor & 1 post-doc
(cid:1) 14 Ph.D. students
Key Projects
(cid:1) 1 SFB/TR-8
(cid:1) 3 European projects
(cid:1) 1 DFG graduate school
(cid:1) 1 BMBF project
(cid:1) 3 projects funded by
industry
Fields of Research
(cid:1) Mobile robots
(cid:1) State and model estimation
(cid:1) Adaptive techniques and learning
(cid:1) Multi-robot coordination
(cid:1) Decision-theoretic approaches
(cid:1) Scene understanding
(cid:1) Manipulation
(cid:1) Autonomous cars
(cid:1) Humanoid robots
(cid:1) Flying vehicles
(cid:1) …
Probabilistic Robotics
What is this Talk About?
SLAM
localization
mapping
integrated
approaches
active
localization
exploration
path planning
A Typical Robot…
proximity senor
wheels with encoders
(provide odometry)
Maps are Important for Robots
(cid:1) Maps model the environment of the robot
(cid:1) Localization is impossible without a map
(“Where am I?”)
(cid:1) Efficient motion planning requires maps
(“How to reach a target location?”)
(cid:1) Reasoning about the state of the world
(cid:1) …
Maps are important for efficiently
solving standard robotic problems
What is “SLAM” ?
(cid:2)
Estimate the pose and the map of a mobile
robot at the same time
poses map observations & movements
Courtesy of Dirk Haehnel
Mapping using Raw Odometry
(cid:2)
Why is SLAM hard? Chicken-or-egg problem:
(cid:2)
a map is needed to localize the robot and
(cid:2)
a pose estimate is needed to build a map
Courtesy of Dirk Haehnel
Particle Filters
Who knows how a particle filter works
?
Explain Particle Filters Skip Explanation
Brief Introduction to Particle Filters
What is a particle filter?
(cid:2)
It is a Bayes filter
(cid:2)
Particle filters are a way to efficiently represent
non-Gaussian distribution
Basic principle
(cid:2)
Set of state hypotheses (“particles”)
(cid:2)
Informally speaking: “survival-of-the-fittest”