Skip to content
Muhan Li

ARXIV 2026

Evolved Dexterity

Zihan Guo*, Shuzhe Zhang*, Muhan Li*, Peiyang Li, Sam Kriegman · *Equal contribution

Robots evolved from scratch learn to rotate, flip, pick and hold objects and to hang on them; the winners are printed and work in the real world without retraining.

MY PART

I built the distributed training and large-scale evaluation pipeline; the largest multitask run trained 11,200 policies from scratch on 16 GPUs in 3 days. I guided the paper.

EVOLVED FROM SCRATCH

Evolution invented this body from scratch, as voxels, to rotate objects in its grip. It is didactyl: two functional appendages, and a third so short that it is vestigial.

BODY:
VOXELSMESHPRINTED
VOXELS
18 795

Every step of the figure

  1. EVOLVED FROM SCRATCH

    Evolution invented this body from scratch, as voxels, to rotate objects in its grip. It is didactyl: two functional appendages, and a third so short that it is vestigial.

  2. SMOOTHED INTO A MESH

    To become a machine, the body first leaves the voxel grid, and its voxels melt into one continuous surface with the same shape.

  3. PRINTED, TESTED IN THE REAL WORLD

    Then the smooth body becomes printed shells with a motor at every joint, and the robot works in the real world. It rotates objects it has never seen almost twice as fast as the best prior design.