Reminds me of Core War. Only that the cells were single memory locations. There are no 'actors' - if you could call it that because cells had no owner; only execution threads have.

Maybe some insights can be derived by comparing the significant experience with Core Wards programs. For example There was no single winning strategy but kind of rock-paper-scissors types of programs. It also allos analysis and exploitation of opponents code - albeit at a very low level.

ADDED: The advantage of this model (which has been used to study evolution of actors) is that it is extremely simple. The disadvantage is that the effect per computation is very high: A simple loop can alter a sizable fraction of the 'world' within short time. Thus no complex analysis of opponents never pays off (except for tests like 'is some opponent at this address').

I like your bot-world as it is kind of a hybrid of other bot worlds and core wars in that it does allow inspection. But I think it's complexity (robots, parts, items, inspection, cells; esp. the winning condition) doesn't cut directly to the problem at hand. The key point missing in core wars is a limit to physical reach. Using a limit to the range of mov-instructions would have made a significant difference - even without going to 2D or 3D (actually that complicates reasoning). Or if mov instructions took more time the further they reach.

I agree that a more gradual winning criteria than dead/alive (as in core wars) would help direct the search for more efficient programs.

See also: The AI Challenge - Ants Article