CoorDex Enables 49-DOF Humanoids to Walk and Grasp Simultaneously
量子位 · wechat · 2026-07-31
A joint team from UNC and UC Berkeley proposed CoorDex, a framework solving the difficulty of simultaneous locomotion and dexterous manipulation in humanoids. The paper was accepted by RSS 2026 Workshop.
Core Challenge
Traditional methods forcing RL to directly control all joints (e.g., 29-DOF body + 20-DOF hand = 49 dimensions) lead to exploration explosion. Robots usually must stop walking to stabilize grasping, as walking alters wrist positions and finger contacts disrupt whole-body balance.
Solution: Motion Priors & Residual Coordination
CoorDex adopts a divide-and-conquer strategy instead of putting all joints into one network:
- Separated Pre-training: Trains distinct "motion priors" for the body and hand, compressing them into low-dimensional latent spaces (16D for body, 12D for hand).
- Residual Tuning: Instead of learning joint angles from scratch, the downstream task policy predicts "residuals" (minor adjustments) for both body and hand based on the current state.
- Coordination Mechanism: A shared coordination trunk understands task phases (approaching, pre-grasping, transporting) before outputting specific corrections, ensuring seamless hand-body cooperation.
Results
CoorDex performed exceptionally well across three continuous mobile manipulation tasks in IsaacLab:
- WalkGrab: 55% success rate, with 0% stopping and 0% falling rates.
- OpenFridge: 66% success rate, averaging a 57.76° door opening angle.
- WalkPickTurn: 89% success rate with minimal orientation error.
Ablation studies show that without this structured division (e.g., using monolithic latent prediction), success rates drop to 0%. The team validated kinematic compatibility on a real Unitree G1. They also introduced AnyBody, an extension allowing coordinated whole-body motion generation from sparse keypoint inputs (e.g., a single wrist).
More from Embodied
- BMW Tests Figure 03 Humanoid Robots in Spartanburg Plant — emmanuelvivier · 2026-07-31
- Samsung Establishes Robotics Division to Accelerate Industrial and Service Bots — emmanuelvivier · 2026-07-31
- HKU's Taku Komura: Building Embodied AI World Models with Consumer Hardware — 量子位 · 2026-07-31
- Physical Intelligence Clarifies Strategy: Solving Physical Intelligence by Any Means Necessary — ZeYanjie · 2026-07-31
- Live Footage: Grinding of a Large Motor Shaft — MatthewChang · 2026-07-31
- M5Stack Workshop: Training Self-Balancing Robots with RL from Sim to Hardware — ShawnHymel · 2026-07-31