SciGroveBeta
Robotics

Any2Any: Efficient Cross-Embodiment Transfer for Humanoid Whole-Body Tracking

Ming Yang, Tao Yu, Feng Li, Hua Chen

Featured May 26, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Making a robot copy human movements is hard and takes tons of computer power; this paper shows a smart way to quickly teach a robot to move like another by first matching their body parts, then making small, targeted adjustments to how it uses its muscles.

In depth
The paper introduces Any2Any, a framework for efficiently transferring pre-trained whole-body tracking (WBT) policies from one humanoid robot to another. It achieves this by first performing kinematic alignment to resolve structural differences in observation and action spaces, then applying dynamics adaptation using lightweight parameter-efficient fine-tuning (PEFT) components to account for robot-specific physical properties. This approach significantly reduces the data and computational cost compared to training from scratch.

Key Takeaways

  • 1
    Any2Any enables efficient transfer of pre-trained whole-body tracking policies across diverse humanoid robots.
  • 2
    The framework decomposes transfer into kinematic alignment (structural mapping) and dynamics adaptation (lightweight fine-tuning).
  • 3
    It achieves competitive or superior tracking performance with only ~1% of the compute and data required for full training, demonstrating significant resource efficiency.

Conceptual Flow

HIGH LEVEL
1
Methodology: How was it done?

The method first makes sure the new robot understands the old robot's body parts, then teaches it how to use its own unique muscles with small, smart adjustments.

Old Robot's Brain
Match Body Parts
New Robot's Brain
Small Muscle Adjustments
2
Results: What did they find?

This new way lets robots learn to move much faster and with less effort than starting from scratch, while still moving just as well.

Start from Scratch
Any2Any Method
Compare Learning
Slow Learning, High Cost
Fast Learning, Low Cost