Professor Noel Sharkey of the Computer Science Department of the University of Sheffield in the UK has been involved in public demonstrations of robotic capabilities for many years.
The news article cited by Joz is tantalizing in its ambiguity and lack of detail. Evidently both predator and prey can experience demise, the predator through starvation. A robot whose batteries go dead is considered deceased, at which point the robot is "reincarnated" with fresh batteries and a copy of the latest program and data of a survivor.
But the article doesn't say precisely what the robots learn. Just saying, "I'm still alive, I must be doing something right" seems insufficient. It would be nice to know how the robots distinguish their own robot species from enemy robot species, what kind of elements comprise their environment, and to what degree they're able to move about in and control their environment. I wonder if one attack is sufficient to "kill" a prey robot, or does it take multiple attacks to achieve this (the latter seems more likely to me). And learning is one thing, while evolving is another. Where, exactly, does the evoution come into play? My guess is that when a successful robot's probram is uploaded that the program is "evolved" before being downloaded into other robots.
At heart this is another example of experiments in evolution through software simulation. Most of the what I've read about this has been purely software in expression, ie, the organisms play in a simulated software environment, sort of like the Sims game. It'll be interesting to hear how this real-world version plays out.
--Percy