When Two Robots Learn 'Secret Codes', Do They Still Recognize Each Other After Changing Bodies?

This research explores the 'reality gap' by training two virtual robots to cooperate in a simplified 2D grid world and then transferring their policies to a 3D physics-based environment. The study finds that while success rates improve with longer time budgets, specific neural weight asymmetries that are negligible in simulation can cause fatal issues like continuous spinning in realistic physical settings. Additionally, social signals evolved primarily function as emergency 'rescue' mechanisms to unstuck robots rather than navigation guides, highlighting the critical importance of reward shaping and environmental fidelity in sim-to-real transfer.