Skip to content
Muhan Li

ICLR 2025 · SPOTLIGHT

Endoskeletal

Muhan Li, Lingji Kong, Sam Kriegman

A generative framework evolves 3D bodies with coupled soft tissue and rigid skeletons, a continuous morphology space, and one morphology-conditioned controller shared across bodies.

SYNTHETIC SPACE OF BODIES

Before it can design a body, the model must learn what a body can be: it studies twenty million synthetic ones, branching skeletons wrapped in soft tissue.

TRAINING SET
20 MILLION

Every step of the figure

  1. SYNTHETIC SPACE OF BODIES

    Before it can design a body, the model must learn what a body can be: it studies twenty million synthetic ones, branching skeletons wrapped in soft tissue.

  2. LATENT GENETIC CODES

    Twenty million shapes are too many to search voxel by voxel, so each body is folded into a short genetic code where nearby codes make similar bodies.

  3. FROM GENOTYPE TO PHENOTYPE, FROM INDIVIDUAL TO POPULATION

    A code is only numbers until a decoder grows it into a body of bone and soft tissue, with about two orders of magnitude more voxels than the earlier freeform robots. A generation of codes becomes a population.

  4. SHARED BRAIN

    No two of these bodies share a skeleton, so rather than train a brain for each, one shared brain reads any skeleton as a graph and drives every joint.

  5. EVOLUTION

    Now that every body can move, each is tested in its own world; those that walk farthest steer the next generation of codes, and the loop begins again.