Osiris

What is Osiris ?

I started off with the idea of making IRIS a basic functional image processing segment for COMRADE. It has evolved beyond its aims to an extent I'd not anticipated. And now, even beyond IRIS-3D (for the moment forgetting the fact that there are lots of refinements still needed), lies Osiris. Osiris is not IRIS, and will soon warrant a much more detailed description than is possible here. Osiris will be the controlling "brain" that IRIS will report to.

Please note that Osiris is still what IRIS was about six months ago, but will evolve as fast as its creator can write code :-) Among the first things it should be able to do is perform localisation. COMRADE's odometry is projected to be notoriously erroneous, so vision-based localisation is a good bet. There are some problems, however.

• Most localisation systems use omnidirectional vision for triangulation purposes or panoramic images. These require mirrors and other fragile equipment which are prohibitively expensive to set up.
• Our team leader (sorry JD, couldn't resist that ;-) wants to run COMRADE on rough terrain (and no, we don't have inclinometers). In which case, odometry is virtually useless, unless there is explicit correction applied, which is also something Osiris should attempt.

This is a classic case of using existing resources to solve an unanticipated problem. I'm currently trying to build a localisation system based on a biological model of place recognition, using neural networks embedded in a self organising map (SOM).

Watch this page for more updates on this exciting approach.

Secondly, and as important, is the issue of constructing 3D maps of the environment during exploration. This task is not difficult at all if we have proper localisation. Turns out, localisation and map-building depend upon each other intimately.

However, since we cannot provide a world map to the robot, the first task has to be localisation. The SOM-based localisation is useful for more high-level tasks, but accuracy in terms of an (x,y) pair is not its intended goal. The problem is currently under active study, and we may use sonar, vision, WLAN or a combination of these tools to achieve robust localisation.