Portland Area Robotics Society

Regular Meeting Minutes

April 5, 2008, 10:30AM 

Agenda:

1.     Welcome

2.  Presentation

            - Introduction to Bayesian Particle Filters

3.  Announcements 

- BotFest ‘08

4.  Update on Open AHRMS

5.      Show and Tell

 

Presentation

Introduction to Bayesian Particle Filters – Professor Barton C. Massey, Portland State University

 

            Authors E-mail - bart@cs.pdx.edu

            Authors Website - http://web.cecs.pdx.edu/~bart/

                                         http://wiki.cs.pdx.edu/forge/bmpf.html

 

1.      What's the problem?

Why do we need to know state of Robot?   Platform Control, Accurate Navigation

 

One Solution - Bayesian Particle Filtering (BPF)

A method of Sensor Fusion – Robot State Space tracking

BPF is the HOT NEW Alternative to Kalman Filtering – Last 10-15 years

 

2.  Approaches

a) Deal with sensor unreliability by adding redundant sensors -

b) Kalman - linear normal distribution estimation of maximum probability position

                        Linear algebra - lots of high level math required

                        Improvements - EKF, UKF

c) Bayesian Particle Filtering - Computational intensive

 

3.  Bayesian Particle Filtering - How does it work?

BPF Inputs

State of vehicle - s

g: dead reckoning

h: measurement function

 

Need to deal with multiple probabilities from Dead Reckoning and sensor measurement.  In order to find robot location you must maximize the product of all state probabilities.  Sounds Simple, Right?   à Space is large

 

Basic Process for BPF:

0) Create copies of vehicle state, Si.  Set all probabilities to 1 (Wi = 1)

1) Propagate each state copy, Si, forward one time of sample.

2) Calculate the probability of each state (Si) given the sensor reading (h) and probability, Wi.

3) Pick and Si to use right now based on maximum probability product, Wi[max]

4) Resample Si.  Use a weighted random selection of states to continue duplication and propagation.

5) Repeat for next time sample.

 

Wi' = h(Si)*Wi [normalize]

 

Bayesian Particle Filter is a method of simulated evolution.  BPF is a useful machine learning technique.

 

Drawback - Computationally intensive - Floating Point can be a plus

Announcements

BotFest '08 Indoor Competition – May 17th at the Portland Children's Museum (Cafe Space)

http://www.portlandrobotics.org/botfest08.php?link_id=26                   

 

BotFest ’08 Outdoor Competition – May 25th at the Oregon Episcopal School (OES)

http://www.portlandrobotics.org/botfest08.php?link_id=26                   

http://www.portlandrobotics.org/PARTS_Outdoor_Challenge.html

 

New PARTS Website – New Content Management System

http://www.portlandrobotics.org                       

           

 

OpenAHRMS Update

 

OpenAHRMS is the Open Source project to develop an Attitude and Heading Reference and Measurement System for Personnel Robotics.  Hardware layout is nearing completion.  The current design will provide USB connectivity to Host controller.

 

http://wiki.cs.pdx.edu/openahrms/

http://svcs.cs.pdx.edu/mailman/listinfo/openahrms

Show and Tell