Skip to content
Muhan Li

ICML 2026

Structured Intelligence

Yibin Wang, Muhan Li, Zihan Guo, Sam Kriegman

Embodiment-conditioned mixtures of control experts co-evolve with latent robot morphologies.

SHARED EXPERTS, A BLEND FOR EACH BODY

One brain for every body plays safe; one for each body learns nothing from the rest. ECo-MoE shares a few experts, and a gate reads each body's genetic code to blend them.

EXPERTS
4
BODIES PER RUN
7 680
BODY
25 / 49
MIX
0.95 · 0.01 · 0.00 · 0.03

Every step of the figure

  1. SHARED EXPERTS, A BLEND FOR EACH BODY

    One brain for every body plays safe; one for each body learns nothing from the rest. ECo-MoE shares a few experts, and a gate reads each body's genetic code to blend them.

  2. EVOLUTION GOES FURTHER

    Over the generations each body comes to lean on one or two experts. Because each body now gets moves that suit it, evolution finds bodies that walk farther than with one shared brain.

  3. TEACH A SKILL, STEER EVOLUTION

    Experts can also be swapped. Plug in one trained on a hand-built four-legged walker, start near its design, and evolution grows bodies with more legs, closer to the demo.

    Experts can also be swapped: here, one trained on a hand-built four-legged walker.

    Started near its design, evolution grows bodies with more legs, closer to the demo.

Previous · Endoskeletal