DPRG
DPRG List  



[DPRG] Re: stella driving down a path

Subject: [DPRG] Re: stella driving down a path
From: Chris Jang cjang at ix.netcom.com
Date: Tue May 15 06:04:49 CDT 2007

>1/10th scale R/C cars are a nice size to handle outdoors.
>
>These guys appear to be navigating visually using a 1/10th
>scale R/C car with a clever omin-directional video camera
>and onboard processing (pdf):
>
><http://infoscience.epfl.ch/getfile.py?recid=63997&mode=best>

An omnidirectional camera would be very nice for purely optical
SLAM. The problem is far more constrained compared to a
conventional monocular camera.

I downloaded the big image (3 megapixel) they put up of the
Kyosho car with the omni-camera. There isn't enough inside for
onboard vision processing given the current state of the art.
It looks like there is an onboard video processor. But as in
the research paper, all of the heavy number crunching is done
on a workstation, mostly in simulation. That's the machine
learning. But it looks like they didn't attempt to miniaturize
everything inside the 1/10 scale car for live operation.

>The Traxxas E-MAXX 4WD 1/10 scale monster truck, which is the
>source of parts for the jBot robot, 
>
><http://www.geology.smu.edu/~dpa-www/robo/jbot/jbot.html>
>
>itself makes a nice platform:
>
><http://www.traxxas.com/products/electric/emaxx/trx_emaxx.htm>
>
>(might need to be geared down a bit - I think top speed of the stock
>E-Maxx is 30+ mph)  I've seen a couple of very nice SRS robo-magellan
>entries based on this platform.  Dave Hylands' website has some good
>pictures:
>
><http://www.davehylands.com/Robotics/Robo-Magellan/>

Traxxas E-Maxx...thanks for this tip. There is a huge variety of hobby
grade vehicles. It is difficult to pick one. That there is a known
good choice makes it easier.

David, you may find this interesting. Sebastian Thrun has a very nice
paper (yet to be published). It is fun to read. The page below has a
link to a PDF.

http://robots.stanford.edu/papers/thrun.stanley05.html

During GPS outages, the localization uses a model that is constrained to
motion only in the direction the vehicle is pointed in. This reminds me
of something you said about jBot. I believe the gyros and magnetometer
give direction. But distance comes from odometry. I think the reason why
jBot does this is similar - too much noise if the localization is not
constrained this way.

More information about the DPRG mailing list