New article on cumulative route improvements


In a new article, I describe a computational model of simplified avian navigation. This model is capable of flying and remembering idiosyncratic routes, and to improve upon them with each new influx of naive individuals. Such inter-generational cumulative improvement has previously been found in pigeons, and this new paper identified the minimal requirements for it to emerge.


Animal researchers were rightfully excited when Sasaki and Biro (2017) showed that consecutive generations of pigeons seemed to display “cumulative culture”. They already knew that pigeons can fly back home upon being released from an unfamiliar location, and that pigeons fly the same route upon being released from that same location several times in a row. What was new in this paper, is that it suggested that pigeons can use collective intelligence to improve their routes over generations!

In the paper, Sasaki and Biro used a diffusion chain to create generational turnover. This means that they started with a single naive pigeon, who made 12 flights home (after being released from the same site each time). At the end of this, it had established a preferred route, and was quite experienced at flying this from memory. Then a naive pigeon was introduced, who had no existing memory of the route. The new pair flew another 12 flights, during which the naive pigeon learned the route. Next, the original (most experienced) pigeon was replaced by a naive one. They flew another 12 flights, after which the most experienced pigeon was again replaced by a naive, and so on for five generations.

By the fifth generation, the pigeons in this experimental condition with generational turnover showed more efficient paths than pigeons who flew in pairs without any change. They concluded that pigeons could have accomplished this by communicating information between each other, and by evaluating outcomes.

Following the definition by Mesoudi and Thornton (2018), this pigeon behaviour meets the core criteria of cumulative culture. This concept is also used to describe technological innovation in humans: we make something new, we teach the next generation how to do it, and then they go on to further improve upon our innovations. Like a ratchet, humanity’s skills improve with each generational step.

It was already surprising to see something akin to cumulative culture in pigeons. However, I wondered if we could take it even further. Could a simple system without the ability to communicate or evaluate what they are doing also show cumulative improvements?

Artificial navigators

I approached this question of “how dumb can we go?” by creating a model of only the core parts of avian navigation. This artificial navigators model comprises four parts:

  1. Goal direction. Agents know roughly where the goal is. This maps onto navigation in pigeons, who know roughly where their home cage is. (They navigate using the sun, a magnetic compass, and/or their sense of smell.)
  2. Social proximity. Agents seek to be near each other, and to align their headings with those who are close. This maps onto flocking behaviour, where proximity to and alignment with other is also sought.
  3. Route memory. Agents learn to recognise visual landmarks along their route, and they fly increasingly precise routes along these landmarks. This is akin to pigeons, who fly idiosyncratic but stable routes.
  4. Continuity. This is a simple trick to ensure that agents don’t make erratic turns. When deciding on where to go next, agents weigh their current heading prominently, so that turns can be smooth. This isn’t necessary for navigation: erratic turns are no problem. However, it does make it more realistic. Pigeons can’t just fully turn around from one moment to the other!

The video below is a visual illustration of the model (on the left) going through several generations. In each generation, a naive bird is introduced. This makes the path more efficient with each generation, as it slowly edges towards the direct line from start to finish.

By turning components of the model on or off, I found that three factors are necessary and sufficient for cumulative improvements in paths to emerge: goal direction, social proximity, and route memory. (This was already my hunch, otherwise I wouldn’t have built the model this way… But you still have to run the simulations to prove it!)

One reason why this is exciting, is that it means that complex communication or thought is not necessary for cumulative improvements occur. In fact, they spontaneously emerge! Stepwise improvements between consecutive generations are a natural consequence of how these agents operate.

Note that the reverse is not necessarily true. That is: pigeons who show cumulative improvements don’t necessarily (only) show it because they align with my model. It could very well be true that they communicate and evaluate, just like Sasaki and Biro proposed. My findings simply say that cumulative improvements are an emergent property in socially navigating systems; not that all socially navigating systems are as simple as my agents!

Read more

The Conversation is a science communication website that uses input from researchers to write accessible articles on their areas of expertise. They published a lovely piece on the robot pigeon findings, and how it relates to their flesh-and-blood counterparts.

You can also read the full paper in the journal PLoS Biology.


  • Dalmaijer, E.S. (2024). Cumulative route improvements spontaneously emerge in artificial navigators even in the absence of sophisticated communication or thought. PLoS Biology, 22(6), e3002644. doi:10.1371/journal.pbio.3002644
  • Mesoudi, A., & Thornton, A. (2018). What is cumulative cultural evolution? Proceedings of the Royal Society B: Biological Sciences, 28520180712. doi:10.1098/rspb.2018.0712
  • Sasaki, T., & Biro, D. (2017). Cumulative culture can emerge from collective intelligence in animal groups. Nature Communications, 8, 15049. doi:10.1038/ncomms15049

Leave a Reply

Your email address will not be published. Required fields are marked *