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

Researchers introduce CoorDex, a learning pipeline that enables humanoid robots to perform complex dexterous manipulation tasks while continuously moving, rather than stopping to grasp objects. The system coordinates high-dimensional body and hand control through latent priors and residual reinforcement learning, demonstrated on a Unitree G1 humanoid with a 20-DOF hand performing tasks like in-motion bottle grasping and fridge operation.

CoorDex represents a meaningful advancement in embodied AI by solving a fundamental limitation in humanoid robotics: the inability to perform precise manipulation during locomotion. Traditional approaches treat loco-manipulation as discrete phases—walk, stop, manipulate, resume—which severely constrains real-world utility. This research bypasses that constraint through a hierarchical learning architecture that decouples body and hand control into separate latent representations while maintaining coordinated execution through shared task context.

The technical innovation addresses a long-standing challenge in high-dimensional control: training stability and sample efficiency when controlling 20+ degrees of freedom simultaneously. By using privileged motion tracking teachers and distilling them into frozen priors, CoorDex reduces the effective action space complexity while preserving natural motion patterns. The ablation studies demonstrate that naive joint-space approaches fail under identical computational budgets, validating the architectural choices.

For the robotics and embodied AI sectors, this work signals progress toward practical humanoid systems capable of household and industrial tasks requiring simultaneous locomotion and manipulation. The demonstrated capabilities—opening doors while walking, grasping while in motion—reflect capabilities essential for autonomous agents in real environments. The approach is generalizable across different humanoid platforms and hand designs, potentially accelerating deployment timelines for companies developing service robots.

Future work likely focuses on sim-to-real transfer reliability, learning from fewer demonstrations, and extending capabilities to higher complexity tasks requiring tool use or bimanual coordination. The underlying latent-prior framework may influence broader approaches to multi-task robot learning.

• →CoorDex enables continuous dexterous manipulation during locomotion, eliminating the stop-and-go limitation in humanoid robotics
• →The system uses latent priors and coordinated residual learning to manage high-dimensional control of 20+ degree-of-freedom hands
• →Experimental validation on Unitree G1 demonstrates practical capabilities including in-motion grasping, door opening, and object manipulation
• →Ablation studies confirm that traditional joint-space control and monolithic latent approaches fail at equivalent computational budgets
• →The modular architecture separates body and hand control while maintaining coordination through shared task context