You have a bit more work to do before deciding on a particular package. The answer to the question will largely be based on what technique you intend to use which can only be determined after having built a database of sample images that indicate the specific aspects of what you want to detect. This field, as you may not be surprised, is largely driven by what content is present within an image.
Once you have created these sample images (which ideally would be exactly what you expect to normally see from the point of view of the robot) you can then better understand what kind of package you are looking for and what skillsets would be required to use it.
At a higher level, it will also depend on your experience with visual processing. If you have someone that's done vision processing before then I'd go with what they are familiar with. If not, RoboRealm is going to be an easier package to experiment with (its got a GUI for every module) and start there. Once you get a better understanding of how things may work, you can have a look at what changes need to be made for deployment. For example, not all robots have powerful PC's onboard, or if 1000's of robots are to be produced that changes what is possible too.
The best package is the one that is going to get the job done ... but you will not really know which one that is until you're done! :-)
Perhaps if you post some images of what you plan to detect we can help further.