I'm interested in self replicating artificial life (with many agents), so after reviewing the literature with the excellent Kinematic Self-Replicating Machines I started looking for software implementations. I understand that the field is still in the early stages and mainly academic, but the status of artificial life software looks rather poor in 2019.

On wikipedia there is this list of software simulators. Going trough the list only ApeSDK, Avida, DigiHive, DOSE, Polyword have been updated in 2019. I did not find a public repo for Biogenesis. ApeSDK, DigiHive and DOSE are single author programs.

All in all I don't see a single very active project with a large community around (I would be happy to have missed something). And this is more surprising considering the big momentum of AI and the proliferation of many ready to use AI tools and libraries.

Why is the status artificial life software so under-developed, when this field looks promising both from a commercial (see manufacturing, mining or space exploration applications) and academic (ecology, biology, human brain and more) perspective? Did the field underdelivered on expectations in past years and got less funding? Did the field hit a theoretical or computational roadblock?

  • $\begingroup$ Why (or how) does artificial like look promising from a commercial point of view? $\endgroup$
    – nbro
    Commented Nov 27, 2019 at 18:03
  • $\begingroup$ It is one of the few ways in which one can sustain a consistent space mining supply chain. Doing the same with individually built instruments would be too expensive and time consuming, while with artificial life you could drop a single probe on a planet to start a mining facility. $\endgroup$
    – Rexcirus
    Commented Nov 27, 2019 at 20:45
  • $\begingroup$ You say "It is one of the few ways in which one can sustain a consistent space mining supply chain.", according to whom? Also, what do you mean by "consistent space mining supply chain"? $\endgroup$
    – nbro
    Commented Nov 27, 2019 at 22:38
  • $\begingroup$ Here you can find many references: en.wikipedia.org/wiki/Self-replicating_machine Plenty of studies have been funded by NASA. By supply chain I mean mining, manufacturing, transportation, refueling and other activities needed to support industrial processes. $\endgroup$
    – Rexcirus
    Commented Nov 27, 2019 at 22:58
  • $\begingroup$ By consistent I mean "of relevant size", so not just few probes going around, but a full industrial scale process, in order to serve space tourism and colonization. $\endgroup$
    – Rexcirus
    Commented Nov 27, 2019 at 23:01

1 Answer 1


I don’t know for certain, but I can make a guess. This is just my opinion, some others may disagree.

The field of ALife has four branches that I’m aware of:

Self-Organizing/self assembly behavior. This is the application you refer to, another context it’s useful is swarm control (for drone swarms, for example). While this is technically ALife, as far as I’m aware it’s not really where most of the emphasis is. Swarm control and self assembly are seen as “different” problems, as machines that can work together and also build more of themselves is interesting (and potentially dangerous), but is missing out on the diversity, the open-endedness that life on earth has. Much of ALife research is focused on trying to formally define this open-endedness and coming up with systems that achieve that. Self assembly and swarm control are interesting and difficult problems, just different. This leads to the other three sides of ALife research:

Coming up with environments, and running tests on them. This is a constant game of coming up with a definition that seems to capture open-endedness, then coming up with ALife sims that meet that criteria but fall short of our expectations. So new definitions are made and we repeat. Geb is a classic example: Geb has passed pretty much every test so far, but it’s fairly uninspiring to watch. Most of those programs you reference chose a particular ALife paradigm, but that paradigm may not be the right one, and is often disappointing. Because we still haven’t found something that really “looks like life”, new paradigms and programs are constantly being created and abandoned when they fail to work (Or perhaps some would have already worked, but the computing time is too much). That’s what you’re seeing. Without any unifying theory or sim that is really convincing, I suspect it’ll stay this way for a while. And because:

  • we still haven’t made much progress since Karl Sims in the 90s, or since Geb (this point is debatable)
  • these sorts of sims don’t really have much commercial use aside from games

the direction of making new simulators seems to be lacking funding and research interest, as far as I can tell. Commercial sim games seem to push the boundary these days.

Fortunately there’s a sub field of cellular automata life that’s pretty interesting, its software is slightly more developed due to the overlap with cellular automata and ease of implementation, and research seems to be progressing there at an okay rate.

Realistically, there seem to be two things people want: novel behavior, and novel bodies. My two cents is that these are separate problems, and achieving both is more expensive than just achieving one. But most of these sims end up not balancing development happening in both of these factors (doing this is very difficult), so one factor develops much further than the other, and this disconnect is disappointing to the sim creator. For example, Geb does behavioural diversity really well, while Karl Sims does body diversity well. Sensitivity to small details like mutation rate or genetic encoding also can be quite frustrating. Fortunately, eventually we’ll sorta get behavioural diversity for free in any sim once RL/AI is really understood well.

The third piece of ALife research I’m aware of is the theoretical side, which right now mostly isn’t really far enough along to warrant practical implementation. One big branch of this is the learning theory side, represented by Valiant’s Evolvability theory and followups. Essentially this talks about what functions are possible to evolve, and using stuff like PAC Learning theory they are able to prove some things. Some of these models are more natural than others, but it’s an interesting perpendicular approach to coming up with sims and seeing if they do what we want. Maybe eventually these two approaches will meet in the middle at some point, but they haven’t yet.

The fourth piece is Artificial Chemistry. I recommend this paper as a somewhat dated overview. While this is technically a field of ALife, and is centered around understanding a chemical system that has the necessary emergent properties, it has broken off into applications that may have industrial relevance. For example, robust self repairing and self assembling electronic systems, DNA computing (DNA is capable of simulating arbitrary chemical reaction networks which are capable of arbitrary computing), and artificial hormone systems for automatic task assignment. This has some software developed, but much of that software isn’t really considered ALife anymore since it has branched off into its own domain.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .