Training whole-body tracking (WBT) models for humanoid robots from scratch demands large-scale data and substantial computation, making rapid deployment on new platforms prohibitively expensive. A new paradigm called Any2Any addresses this challenge by enabling efficient cross-embodiment transfer, according to a paper on arXiv. Any2Any allows an existing WBT specialist to be reused on a new humanoid robot with only a fraction of the original training cost.
The High Cost of Embodiment-Specific Training
Whole-body tracking models are foundational for humanoid robots to imitate diverse motions with high fidelity. Training such models from scratch for each new robot embodiment is costly in both data collection and compute resources. The research team behind Any2Any questioned whether pretrained WBT models could transfer across embodiments with minimal adaptation, leading to the development of their new framework.
Any2Any: Kinematic Alignment and Dynamics Adaptation
Any2Any proceeds in two main steps. First, kinematic alignment aligns the input and output spaces of the source and target humanoids, ensuring the pretrained source policy can be meaningfully reused on the target embodiment. Second, dynamics adaptation applies lightweight parameter-efficient fine-tuning (PEFT) components to selected dynamics-sensitive modules. This preserves useful behavioral priors while enabling targeted adaptation to the new robot.
Experimental Validation Across Platforms
The researchers conducted extensive experiments on multiple humanoid platforms and pretrained backbones. Notably, Any2Any was used to transfer Sonic models pre-trained on Unitree G1 to LimX Oli and LimX Luna. The results show that Any2Any substantially accelerates convergence and reduces training cost compared with training from scratch, while achieving competitive or superior tracking performance. A key finding: using only 1% of the compute and data required for full training, Any2Any successfully completed the transfers.
| Aspect | Training from Scratch | Any2Any Transfer |
|---|---|---|
| Data required | 100% | 1% |
| Compute required | 100% | 1% |
| Tracking performance | Baseline | Competitive or superior |
| Convergence speed | Slow | Substantially accelerated |
The experiments spanned multiple humanoid platforms and pretrained backbones, demonstrating the generalizability of the approach.
A Scalable Path Forward
The results suggest that pretrained WBT specialists can be efficiently reused across embodiments, providing a scalable path toward deploying humanoid whole-body control on new robots. This could accelerate the development and deployment of humanoid robots in various applications by reducing the cost of adapting motion control software to different hardware platforms.
More results and videos are available on the project page linked in the paper. The authors of the study are Yang, Ming, Yu, Tao, Li, Feng, Chen, and Hua.
