CoEvolve, introduced in a preprint on arXiv, treats multi-robot coordination not as a fixed plan but as executable code that can evolve. The framework builds strategies from a robot's existing motor skills, combining motion primitives, semantic objectives, temporary roles, and bounded execution parameters into a program that defines how the team should act.
A central component is Verifier Guided Coordination Evolution. Each executed transition is physically verified; if a behavior is rejected, the system revises the coordination from the last accepted checkpoint rather than restarting. Verified experiences are retained and used to guide later decisions, allowing the system to improve over time.
The authors evaluated CoEvolve with two simulated Unitree G1 humanoids carrying a rigid payload through constrained corridors. Across 15 base task runs with three multimodal backbone families, the framework completed 8 tasks, and every backbone achieved end-to-end success. Additional tests showed verified placement on two new corridor layouts and successful initialization from both same-backbone and cross-backbone experience.
However, the paper also reports persistent limitations in motion precision and contact stability. The authors argue that executable coordination is a practical interface for adapting long-horizon multi-robot collaboration under shared physical constraints, but the results suggest that low-level control quality remains a bottleneck.