ECo-MoE: Modular AI brains let robots evolve more efficiently

A team led by Sam Kriegman at the University of Notre Dame has published ECo-MoE (Embodiment-Conditioned Mixture of Experts), a framework that bridges two extremes in robot co-design: training a separate control policy for every individual robot (inefficient) or using a single monolithic controller for all robots (overly conservative). ECo-MoE maintains a unified library of neural expert modules that are dynamically gated based on the latent coordinates of each robot's physical structure. This means different body plans activate different combinations of learned sensorimotor circuits, allowing one part of the controller to be updated without breaking others.

The approach also introduces a novel capability called 'evo by demo,' where pretrained expert policies can be injected directly into the mixture. These pre-trained modules steer freeform evolution toward unexplored regions of the design space that contain desired morphological traits, effectively guiding the evolutionary algorithm toward canonical structures defined by the experts. The result is a scalable, modular system that preserves hard-earned knowledge across generations while maintaining the flexibility to adapt to new species of robots. The team has released code and videos demonstrating the method.