Behavioral Persistence and Incomplete Functional Transfer of Co-evolved Communication in Evolutionary Robotics

Behavioral Persistence and Incomplete Functional Transfer of Co-evolved Communication in Evolutionary Robotics. By Fernando Montes-Gonzalez, from the Universidad Veracruzana. Published on the twenty-ninth of September. Its code repository is public but slight: seven source files.

Two robots evolved a way of signalling to each other in a 2D simulation. The question here is what happens when that communication protocol is moved directly into a 3D physical environment, with the network weights left exactly as they were and no retraining at all. The answer the study reports is that the behaviour persists, but its function does not fully carry over. The transfer is partial, and it is asymmetric between the two robots: one of them manages a little, and the other manages nothing. The title captures both halves, behavioural persistence on one side and incomplete functional transfer on the other.

The setup uses two robots of the e-puck type. Each is controlled by a GRU network with a residual connection, and the pair is tested on a task of seeking food while signalling socially. The protocol was co-evolved, meaning the two agents arrived at their signalling together. Getting stable physical operation took three corrections to the layer that translates between the network and the robot's sensors and motors. One of these was calibrating a hunger term, and that calibration was guided by a measurable asymmetry in the trained residual weights.

Success was defined strictly, as both agents reaching the food area. Even after the corrections, one agent reached the food source in one of thirty tested seeds, and the other did not reach it in any of them. A further experiment added explicit directional information to the social channel, so that the signal said more about where to go. That produced observable changes in the trajectory of the receiving agent, and improvements in several specific cases. It was still not enough for the second agent to reach the food.

The scale of the evidence is small: two robots, one task, thirty seeds. The interpretation is offered with hedges. The author suggests that the shortfall may not be explained by signal translation alone, but also by the agents' ability to navigate under the new physical constraints. That is presented as a suggestion, not as an isolated cause. And the three corrections to the translation layer mean the physical system is not identical to the simulated one, so the test is of transfer with adjustments rather than transfer untouched.

The broader lesson the author draws is about what has to travel with a protocol. Carrying the signals across may not be sufficient on its own. Successful transfer of emergent communication may also depend on preserving the ecological and navigational conditions in which that protocol first evolved. A language that grew up in one world may need some of that world around it in order to keep doing its job.