K2

K² · Artificial intelligence

A humanoid that grabs bottles without breaking stride

The authors introduce CoorDex, a learning pipeline that converts high-dimensional body and dexterous hand control into coordinated latent residual control for dexterous loco-manipulation on the move.

SL
SL

Sikai Li, Shuning Li, Zhenyu Wei et al.

6 authors · cs.RO, cs.AI, cs.LG

arXiv preprintArtificial intelligenceJun 2026 · ~70s read

The 30-second scan

Most robots handle objects like a clumsy waiter: walk to the table, stop, pick up the cup, then start walking again.

  1. CoorDex starts from simulated whole-body and hand demonstrations, trains privileged motion tracking teachers for the humanoid body and dexterous hand, and distills them into proprioception-conditioned latent priors used as the action space for downstream residual reinforcement learning.
  2. A coordinated latent residual policy composes the priors through shared task context and separate body-hand residual heads, preserving natural whole-body motion while improving finger-level contact reliability.
  3. CoorDex enables a Unitree G1 humanoid with a 20-DoF WUJI hand to execute dexterous manipulation while in motion, including non-stop bottle grasping and carrying, fridge door opening on the move, and cube pick-and-turn.
20-DoF WUJI hand on a Unitree G1 humanoidAblations show joint-space PPO, joint-space hand control, and monolithic latent prediction all fail under the same reward budget