|
[DPRG] Are representations what makes Can Can hard? and Odometry
Subject: [DPRG] Are representations what makes Can Can hard? and Odometry
From: Chris Jang
cjang at ix.netcom.com
Date: Fri Dec 8 10:50:38 CST 2006
> That is, I wondered why is it not sufficient to do
>navigation based on sensor input that is constantly updated. It
>wasn't until DPA was talking to Will about calibrating odometry
>that I read a good-enough reason for doing odometry: Researchers
>concluded that it is far easier to get good behavior using
>odometry than it is by trying to use representation and sensors.
This is very true. Perception of the environment is compute intensive.
Embedded computers are only just now starting to reach performance
levels that can do it. We are limited by the technology we have.
The history was the same in other areas. Digital flight control systems
became possible after the Apollo program because of advances in
embeddable computers. Digital photography was possible several
generations ago but only recently has the cost dropped to the point
where it has displaced film.
When I started down the path of computer vision, I had no idea how
difficult it is. In my opinion, it is right on the edge of available technology.
So it is possible but only just barely. This is why Nikon and Fuji can put
face detection focus modes in point and shoot cameras. In contrast,
odometry and inertial guidance are mature technologies with low
computational requirements. That has a design spillover of lower power
requirements and smaller computing packages (microcontrollers) which
allow more flexibility in robot design.
Many people with the DPRG are using SONAR. This can also be expensive.
In principle, pulsed time of flight should be easy to interpret as distance.
But in practice (as others have related - I have no experience with SONAR),
there are reflections, echos and noise. Where is the real signal? Non-trivial
computation may be required to isolate it.
More information about the DPRG mailing list
|
