The value of information and the emergence of communication in evolved e-puck robots

The value of information and the emergence of communication in evolved e-puck robots. By Fernando Montes-Gonzalez, from the Universidad Veracruzana. Published on the ninth of October. Its code repository is public and real: about thirty source files.

When does communication arise between autonomous agents, and what part does an innate architecture for choosing actions play? The answer offered here is that communication emerges in proportion to the value of the information it carries, and that innate scaffolding designed for an agent on its own can get in the way when the task is social. Robots signal when knowing what the other robot knows has real value, and they stop when it does not. Built-in behaviour that suits a lone robot can become a burden once two robots have to cooperate.

The setup uses two e-puck robots, reconstructed in PyBullet from their Webots model and validated against it. Together they must find a feeding zone that a robot can detect only by passing over it. Each robot can emit a one-bit signal, and the receiver detects which direction it came from. Control comes from a GRU recurrent network with a residual path. Sometimes that network works alone; sometimes it is one channel inside an Embedded Action Selection Architecture, called EASA. Fitness counts only the time the robots spend over the food. Nothing rewards signalling directly.

Give the robots a compass pointing at the food, and use of the signal disappears completely. Without a compass, in a small arena, three independent runs reached similar performance by different routes. One used a referential signal, one used a permanent beacon, and one did not use the signal at all. When the arena is expanded, the strategy built on a referential signal generalizes best. Evolving directly in the large arena does not make communication more frequent, but it does pull the strategies apart, and there the referential signal clearly beats searching alone.

The second strand concerns the controller itself. Inside EASA, the innate reflexes constrain the network and lower performance. As the architecture's priorities evolve, evolution consistently weakens those reflexes. The scaffolding that was meant to help is, in this social task, something the evolutionary process works to undo.

The evidence is narrow by design. There are two robots, one foraging task, and a signal of a single bit. The small-arena result rests on three runs, which landed on three different strategies, so it shows what is possible rather than what is typical. The conclusions are phrased as suggestions, and they concern this simulated setting, with robots rebuilt in PyBullet, rather than communication in general.

Read together, the results trace a simple economy. Where information is free, as with a compass, nobody signals. Where it is scarce and the space is large, signalling about the food pays and wins out. And reflexes built in for a solitary agent were something evolution chose to loosen once a partner entered the picture.