Any2Any: Efficient Cross-Embodiment Transfer for Humanoid Whole-Body Tracking
The Any2Any framework introduces a novel method for the effective transfer of pretrained whole-body tracking (WBT) models among various humanoid robot designs, requiring minimal data and computational resources. Detailed in arXiv paper 2605.23733, Any2Any begins with kinematic alignment between the source and target humanoids to synchronize their input and output spaces, enabling the reuse of the source policy. Subsequently, it utilizes lightweight parameter-efficient fine-tuning (PEFT) for adapting dynamics-sensitive components. This strategy mitigates the significant expense associated with training WBT models from the ground up for each new platform, thereby promoting swift deployment.
Key facts
- Any2Any is a paradigm for cross-embodiment transfer of whole-body tracking models.
- It performs kinematic alignment between source and target humanoids.
- It uses lightweight parameter-efficient fine-tuning (PEFT) for dynamics adaptation.
- The approach reduces the need for large-scale data and computation for each new humanoid platform.
- The paper is available on arXiv with ID 2605.23733.
- Whole-body tracking models enable humanoid robots to imitate diverse motions with high fidelity.
- Training WBT models from scratch is costly and data-intensive.
- Any2Any enables rapid deployment on new humanoid embodiments.
Entities
Institutions
- arXiv