CoorDex: Coordinating Body and Hand Priors for Continuous Dexterous Humanoid Loco-Manipulation

2026-06-22 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

CoorDex introduces a novel learning pipeline for continuous, high degree-of-freedom (DoF) humanoid loco-manipulation. It addresses the common simplification of stop-and-go processes and low-DoF end effectors. This system enables a Unitree G1 humanoid, equipped with a 20-DoF WUJI hand, to perform dexterous tasks while in motion. Examples include non-stop bottle grasping and carrying, fridge door opening, and cube pick-and-turn. The pipeline begins with simulated whole-body and hand demonstrations, training privileged motion tracking teachers. These teachers are then distilled into proprioception-conditioned latent priors. These priors serve as the action space for downstream residual reinforcement learning. A coordinated latent residual policy integrates these priors via shared task context and separate body-hand residual heads. This maintains natural whole-body movement and improves finger-level contact reliability. Ablation studies confirm that this latent-prior interface and coordinated residual structure are crucial for training high-dimensional, contact-rich loco-manipulation. Traditional joint-space PPO or monolithic latent prediction methods fail under the same reward budget.

Key takeaway

For Robotics Engineers developing humanoid systems, CoorDex offers a robust method to overcome the limitations of traditional stop-and-go manipulation. You should consider adopting its latent residual control pipeline to integrate high-DoF dexterous hand control with whole-body locomotion. This approach enables more fluid, continuous, and complex humanoid tasks, improving contact reliability and overall system performance in dynamic environments.

Key insights

CoorDex enables continuous, high-DoF humanoid loco-manipulation by coordinating body and dexterous hand control through latent residual learning.

Principles

• Latent priors improve high-dimensional contact-rich manipulation.
• Coordinated residual structure enhances finger-level reliability.
• Distilling teachers into priors optimizes action space.

Method

CoorDex trains privileged motion tracking teachers, distills them into proprioception-conditioned latent priors, and uses these frozen priors as the action space for downstream residual reinforcement learning.

In practice

• Grasp and carry objects without stopping locomotion.
• Open doors while moving.
• Perform complex object manipulation like pick-and-turn.

Topics

• Humanoid Robotics
• Dexterous Manipulation
• Loco-Manipulation
• Reinforcement Learning
• Latent Space Control
• Whole-Body Control
• Unitree G1

Code references

• physical-superintelligence-lab/humdex

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• ROVE: Unlocking Human Interventions for Humanoid Manipulation via Reinforcement Learning — ROVE improves humanoid VLA manipulation by using Optimistic Value Estimation and cross-embodiment videos to learn from imperfect human interventions.
• HANDOFF: Humanoid Agentic Task-Space Whole-Body Control via Distilled Complementary Teachers — HANDOFF uses distilled multi-teacher control for robust, intuitive humanoid whole-body command, enabling diverse manipulation and locomotion.
• ContactExplorer: Contact Coverage-Guided Exploration for General-Purpose Dexterous Manipulation — CCGE leverages state-conditioned contact coverage to enable efficient, general-purpose exploration in dexterous manipulation.
• Proprioceptive-visual correspondence enables self-other distinction in humanoid robots — Proprioceptive-visual correspondence enables robots to distinguish self from others and learn a 3D self-model without prior knowledge.
• Mind Your Steps: A General Learning Framework for Accurate Humanoid Foothold Tracking — The framework enables terrain-agnostic 3D humanoid foothold tracking by dynamically sampling goals and using a robust target representation for real-world transfer.
• Learning New Tasks via Reusable Skills: Skill-Compositional Experts for Embodied Continual Learning — SCE enables robots to learn new tasks continually by composing reusable skills, mitigating catastrophic forgetting in embodied learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.