This question presents a lot of opportunities for further investigation with this kind of model. In our system as it currently stands, we are modeling a situation in which there is no explicit harm caused by the collisions, but we could use the same system to investigate this question directly.

From an evolutionary point of view, there could be interesting trade-offs between the complexity of social sensing and organization, the potential harm caused by collisions, and the benefits of organized behavior. We could model evolutionary hypotheses with this system using the approach outlined by (Long, 2007). This is complete speculation, but one could imagine an evolutionary trajectory where passive dynamics and the benefits of organized behavior put pressure on the behavior of individuals to exploit the passive dynamics, creating the kind of system we have modeled with our robots. Then the harm of passively organizing imposes pressure to evolve social mechanisms to accomplish the same organization without collisions. In this kind of a system, the harmfulness of collisions could be treated as a variable. A likely hypothesis is that if the harmfulness of the collisions outweigh the benefits of the organized behavior, the system will either find another way to generate organized behavior or will avoid it.

Reference:
Long, J.H. Jr. (2007). Biomimetic robotics: building autonomous, physical models to test biological hypotheses. Proceedings of the Institution of Mechanical Engineers, Part C, Journal of Mechanical Engineering Science. 221, 1193-1200.

The common goal functions as an attractor, bringing all of the agents together so that the passive dynamics of self-organization will cause them to align. Whether or not a joint goal would achieve similar outcomes probably would depend on what exactly the joint goal was and if it fulfilled the role of an attractor.

In social models of this phenomenon (e.g. Reynolds 1987), the goals of individual agents are also like the “common goal” scenario we’ve modeled here. Usually each agent is attracted to other agents in the group, which depends on some ability to detect neighboring members of the group, but doesn’t require any explicit representation of a joint goal among members of the group. Put another way – the group can coordinate if every agent in the group has the goal “move towards other agents in the group” (a common goal), and it is not necessary for the collective group to have some notion of “make the group as condensed as possible” (a joint goal). Nevertheless, in an evolutionary sense, the individual reward does depend on the group coordination, much like a joint goal! Coordinated behavior has many benefits (and costs) for individuals in the group (Sumpter, 2010) but these depend on the ability of the group to successfully coordinate.

References:

Reynolds, C. W. (1987) Flocks, Herds, and Schools: A Distributed Behavioral Model, in Computer Graphics, 21(4) (SIGGRAPH ’87 Conference Proceedings) pages 25-34.

Sumpter, D.J.T. (2010). Collective Animal Behavior. Princeton University Press.

Thanks for the questions!

(1) Yes, absolutely. The physical collisions between the robots is what causes them to align. There are nice formal models of this kind of interaction that apply to the system we created (see section 4.2.1 of Vicsek & Zafeiris, 2012). When we manipulate how goal-directed the robots are, we are affecting the probability of collisions. If the robots all share a goal, they congregate in the same area of the tank, producing lots of collisions that cause the robots to align.

(2) We would expect the same outcome in three-dimensions. Simulations of social interactions work well in three-dimensions (e.g. Reynolds 1987), and despite the fact that our model achieves coordination without social interaction, the underlying mechanisms in both cases are quite similar. Social models often involve three elements, attraction to nearby group members, avoidance of extremely close group members, and alignment with other group members. Our model can be interpreted using the same ideas: attraction is produced by the common goal, and avoidance and alignment are generated by physical interactions.

(3) There is empirical evidence that fish use social information to school (e.g. Partridge & Pitcher, 1980), and we certainly don’t want to claim that our model suggests that animals coordinate in an entirely asocial manner. Rather, we think that our results show that asocial mechanisms can produce similar results to social ones, and that it is therefore important to consider the role that asocial mechanisms play in generating coordinated behavior. There are certainly advantages to social mechanisms (no potentially harmful collisions, longer-range modification of behavior) that real animals would evolve to take advantage of.

References:

Partridge, B. L., & Pitcher, T. J. (1980). The sensory basis of fish schools: relative roles of lateral line and vision. Journal of Comparative Physiology,135(4), 315-325.

Reynolds, C. W. (1987) Flocks, Herds, and Schools: A Distributed Behavioral Model, in Computer Graphics, 21(4) (SIGGRAPH ’87 Conference Proceedings) pages 25-34.

Vicsek, T., & Zafeiris, A. (2012). Collective motion. Physics Reports.

Our model suggests that when similar organisms share the same asocial goal, coordinated behavior can emerge with the right physical dynamics. There are several models that show similar results with social mechanisms/goals (e.g. Reynolds 1987). Therefore, it seems like whether the goal is asocial or social doesn’t particularly matter in terms of what kind of behavior can be created.

One potential difference between asocial and social goals is that asocial goals are, by definition, not directly related to the coordination of the group. Our system models a behavior like foraging at a food patch, which is important for the fitness of the agent independently of the coordination that it creates. However, some social goals may be directly relevant to coordination. If an agent has a goal to be near conspecifics and travelling in the same direction as its neighbors, then coordination will result (Reynolds 1987). This has potentially important implications for understanding the evolution of these coordinated behaviors. Since coordinated behavior has benefits and costs for individuals in the group (Sumpter, 2010), goals that produce coordinated behavior could be under selection pressure depending on the relative cost and benefits of organized behavior and the individual goals that generate it. For asocial goals that generate coordination, the costs and benefits of the coordination need to be balanced with the outcome that the goal is actually trying to achieve (such as feeding at a food patch). However, social goals that are entirely related to the coordination of the group only contribute fitness to the agent to the extent that the coordinated behavior improves fitness.

References:

Reynolds, C. W. (1987) Flocks, Herds, and Schools: A Distributed Behavioral Model, in Computer Graphics, 21(4) (SIGGRAPH ’87 Conference Proceedings) pages 25-34.

Sumpter, D.J.T. (2010). Collective Animal Behavior. Princeton University Press.

Great question. This is the kind of question that is easier to answer with a computational model than a physical embodied system since exploring a parameter space is much faster in simulation. Luckily, someone else has done this! Grossman, Aranson, and Ben Jacob (2008) built a computational model that closely mirrors the physical system we constructed. They showed that it can exhibit several kinds of collective motion patterns, including circular milling, but also swarm migration where an entire group moves together in a particular direction, and complex patterns that involve multiple semi-stable structures.

A crucial factor in the model is density. When the group density is high enough, the agents self-organize. Our robots achieve a dense clustering by sharing a common goal which causes them to all try and occupt the same area of the tank. The model of Grossman et al. achieves the same outcome by simply have a dense group from the start, constrained by the boundaries of the simulated world. In their model, boundary shape changes the dynamics of the group, which suggests that we could achieve different patterns of organized behavior in our physical robots by programming in different asocial behaviors. For example, having all robots track a moving target would likely produce a swarm migration pattern.

Social sensing could be thought of as a way to mitigate the density requirement. With social sensing, reaction to a neighbor can happen at a distance. This makes it possible to coordinate at much lower group densities, which is obviously very important for biological agents that do not want to collide with other members of their group.

Reference:

Grossman, D., Aranson, I.S., and Ben Jacob, E. (2008). Emergence of agent swarm migration and vortex formation through inelastic collisions. New Journal of Physics, 10.

Thanks Josh. Okay. So this seems highly intuitive to me. Why do we need simulations and autonomous robots to tell us that having a large number of individuals attempt the same thing in close proximity will result in unique group behavior (such as following the moving target?). How is this useful?

It is obvious that programming the robots to swim towards the light will cause them to congregate around the light. That would be a very boring result! But the coordination that they exhibit, which hopefully you can see in the videos of them swimming, goes beyond simply following their programmed behavior. The robots exhibit very stable behavior as a group – meaning that they move as a cohesive unit – despite not being programmed to do this in any explicit way. With our quantitative analysis of this behavior, we can see that the coordination occurs well after the robots congregate around the light, so simply being programmed to swim towards the light is not enough to produce the coordination we observe. We think this is an interesting result on its own, in that it extends the predictions of the model I referenced above towards the biological context of autonomous goals. Additionally, now that we have established that coordinated behavior can occur as a result of common goals with this robotic system, we can now move onto questions about the origins of social behavior, which I’ve touched on a bit in my responses to the other judges’ questions.

Hi, Rob. It’s certainly true that there are many ways to get coordinated behavior through passive dynamics (another great example is Tarcai et al. (2011): http://iopscience.iop.org/1742-5468/2011/04/P04...). We like the embodied and embedded aspect of the robots because we can show that coordination is still achieved despite the noisy aspects of having real autonomous agents.

I’m not sure if the idea that exclusively asocial factors can generate coordinated behavior was in doubt, but models that make this assumption seem to be relatively recent (see section 4.2.1 of Vicsek and Zafeiris (2009): http://arxiv.org/pdf/1010.5017v2.pdf), and physical demonstrations even more so.

Yes, this is certainly a point in favor of computational models as opposed to physical systems. We did not specifically test variations in steering ability or program – mostly because the behavior we were after emerged immediately in the system we built and we didn’t need to tinker with parameters to generate coordination. This could either speak to the robustness of the phenomenon or a stroke of luck on our part.

Right! This is explicitly asocial behavior. We are contrasting this with models like Reynolds’ Boids which use social factors to achieve similar outcomes.

Hi Manny! Think of the robots like bumper cars. Imagine everyone driving the bumper cars are going around the track in the same direction. Now, imagine you are driving a bumper car, but you are trying to go around the track the opposite direction from everyone else. This will be really tough because you’ll keep running into people. But, if you go around the track in the same direction as everyone else, the ride will be smooth and even if you hit someone else it won’t be a jarring collision. Less fun, perhaps! The robots end up in the situation where everyone is going around the track in the same direction because if a few robots try to go the “wrong” direction, then the other robots bump them back into alignment.

I take a comparison to Braitenberg vehicles as a high compliment. Thanks!

You are right on track with the problem you’ve highlighted. We think our system can be used to explore questions about the origins of social behavior. We can set up artificial selection experiments with these robots. One implication of the finding is that you don’t need selection for grouping to get these asocial groups, since the mechanism that produces the grouping is useful in other contexts. But, once you have grouping like this, with the benefits and costs that are associated with it, then selection might favor social sensing as a means to make the grouping more efficient.

Thank you!

The milling fish example is related in the sense that the model produces similar behavior at a group level. I think it is highly unlikely that the underlying mechanisms are the same.

I think the agility/maneuverability question could certainly be tested with a model. If it’s right, I’m not sure that it would limit the application of the model to tugboats and bumper cars. There might be other biological organisms that would fit the requirements found by a model of the behavior.

Can you give us an example of a behaviour or organism that would fit your model?

Some single-cell organisms seem to match this model pretty closely (Rappel et al., 1999): http://srnano.ucsd.edu/~rappel/pub/vortex.pdf

Since the asocial mechanisms on its own is relatively limited in terms of the behaviors it can produce (based on the fact that the agents must be in close proximity) and has costs that animals might want to avoid (also based on the fact that the agents must be in close proximity), then it makes sense for social mechanisms to be favored. However, since the asocial mechanisms can create grouping without selection for grouping, it suggests an evolutionary hypothesis about the origins of social mechanisms. We could test a version of this hypothesis with this model, by allowing the robots to evolve social sensing.

How would you test your evolutionary hypothesis with these robots? You say, by allowing them to evolve social sensing, but how would this be done?