HoloMotion-1: Humanoid Motion Foundation Model for Zero-Shot Whole-Body Tracking
The technical document presents HoloMotion-1, a foundational model for humanoid motion that enables zero-shot tracking of full-body movements. Its primary advancement lies in enhancing control-policy training through a vast hybrid motion dataset. Dominant motion diversity is derived from video-reconstructed actions captured in real-world scenarios, while curated motion-capture and proprietary motion data ensure superior supervision and practical application. This approach transcends traditional MoCap-only training, allowing the policy to learn from a wider range of behaviors, capture environments, and movement styles. However, challenges arise from reconstruction noise, mismatches between source domains, inconsistent motion quality, and temporal modeling amid significant behavioral variations. To tackle these issues, HoloMotion-1 employs a robust temporal modeling system and a sparse architecture.
Key facts
- HoloMotion-1 is a humanoid motion foundation model for zero-shot whole-body motion tracking.
- It uses a large-scale hybrid motion corpus for training.
- Video-reconstructed motions from in-the-wild videos provide dominant motion diversity.
- Curated motion-capture and in-house motion data provide higher-fidelity supervision.
- The model moves beyond conventional MoCap-only training.
- Challenges include reconstruction noise, source-domain mismatch, uneven motion quality, and temporal modeling.
- HoloMotion-1 integrates large-capacity temporal modeling and a sparse architecture.
- The report is published on arXiv with ID 2605.15336.
Entities
Institutions
- arXiv