Date: Wed, 10 May 2006 02:14:38 -0500 (CDT)
Subject: robot is a platform for the vision system
X-UID: 170

Any significant value from this project will be the vision system. This is 
the hard problem. And it's the new long term objective.

1. reliable terrain mapping using optical cameras on inexpensive hardware
    (i.e. webcams with low power single board computers)

Most other issues are relatively easy in the sense that there's no basic 
research required. This project has only appeared difficult so far because 
I did it the wrong way. I built too much and did not buy off-the-shelf.

After posting the first cut drive02 vision system movie to the DPRG 
mailing list, there was some valuable feedback. For me, this clarified 
what is going on and where this project is going.

The limiting factor for many autonomous robots is that they can not see. 
Supporting video requires more computational performance which means 
higher power, bigger batteries, more weight, etc. But mostly the limiting 
factor is software. Developing this kind of software is very expensive. 
I'm running into this now. The only way to know if the image processing 
code works is to test it with real life data sets. I may have to acquire 
more computing power, collect a large number of data sets with 
corresponding ground-truth maps, and then use optimization methods to 
search for filters that work. In effect, this is what I'm doing with a 
manual process at a slow rate.

There are many published ideas for autonomous robot vision systems. There 
is not a lot of source code available to the public. I think one reason is 
the perceived value in keeping technology secret for competitive research 
advantage or commercialization. Another reason is the "demo effect". 
Often, the technology is very brittle. It fails in embarrassing ways 
despite appearing very impressive under the right conditions. I know that 
my stuff is like this.

Even vision software with severe known limitations would be useful. If it 
were easy enough to use, video camera input and occupancy grid terrain map 
output, this would be enough to allow many robots to do interesting things 
under the right conditions. The problem today is that this software does 
not even exist so far as I am aware.

Anyway, I want to get the robot driving autonomously over a short length 
of path surrounded by grass in the short term. But I see this as one step 
along the way to developing optical camera based robot vision software. 
From the start, I knew this robot would serve as a development platform to 
gain experience and write software for the next robot.