From the demo video I saw (not a live demo), a table full of typical tools was flashed on the screen (overhead view), then right after a count of the different tools was displayed. From what I understand (and reading Greg's description above) the "shapes" (and colors and patterns) are programmed into the code somehow, then the code and/or hardware (I don't know which) does the dirty work grinding out which is what. Fairly quickly, I might add. Thus "shapes" could be tool profiles or letter outlines or game pieces. Same with colors. Size and rotation didn't matter, but the profile had to be similar to the pre-programmed profile. In a word----
I guess for 2009 the signs "Turn Left!" "Go Fast!" could really be read by the robot.
I'm thinking there must be a catch, or a trade-off, or something. The problems with the CMUcam was not that it couldn't read green vs red vs blue -- it could -- but that it would pick up arena lights from across the field. (I'm looking at you, Boston's BU arena!) The shapes I'm thinking have to be directly in front of the camera and on a plain contrasting background to be picked up clearly. I can't wait to see how well it really works.
Daniel, my guess is that the software/hardware can determine shape (profile), but
your software would have to save size/rotation of the shape each "frame" to track differences between frames to determine trajectory. Or, wait and see.