# Use of virtual worlds (e.g. Second Life) for training Artificial General Intelligence agents?

There is emerging effort for Third Wave Artificial Intelligence (Artificial General Intelligence) (http://hlc.doc.ic.ac.uk/3AI_HLC_2019.html and https://www.darpa.mil/work-with-us/ai-next-campaign) and it covers the open-ended life-long machine learning as well. Currently machine learning agents are being run on quite immutable and simple games like Atari and Go. But what about the efforts to build and run machine learning adaptable agents (or even teams of them) in the virtual worlds (like Second Life) which are complex, expanding and in which the interaction with human representatives happens? Are there efforts to do that?

I have found some articles from 2005-2009, but Google gives no recent literature on queries like Reinforcement Learning Second Life etc.

So - maybe there are some efforts to do this, but I can not just Google it.

My question is - are there references for machine learning agents for virtual worlds and if not - what are the obstacles for trying to build them? There are little risks or costs for building them for virtual worlds?

https://meta-guide.com/embodiment/secondlife-npc-artificial-intelligence is some bibliography and it is lacking recent research, for example.

• To clarify, you're asking about agents that can "live" inside a virtual world and not necessarily about any research on artificial general intelligence, right?
– nbro
Sep 7 '20 at 22:24
• I have spent several days looking exactly at the same question. First, let me say that I am very far from this field. It seems that now embodied AI research shifted from game-like virtual worlds to small indoor simulated environments. "A survey of embodied AI: from simulators to research tasks" is an example with plenty of references. So it appears that now this kind of research has more applied nature, eyeing deployment to household assistant robots. Jun 6 at 2:39
• The only example of the kind of research I was looking for (and apparently @TomR too) and was able to find was Microsoft Project Malmo, but it does not seem very active. Jun 6 at 2:41
• Similarly to @TomR, I am kind of puzzled by the lack of this type of research. For me as a non-expert, it seems that such embodied AI research using open-ended virtual worlds, populated with human players, could be one of the few currently realistic ways to get actually intelligent AI in near future. I would appreciate response from somebody more knowledgeable than me. Jun 6 at 2:47