OmniHumanoid: Streaming Cross-Embodiment Video Generation with Paired-Free Adaptation
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 1 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Yiren Song et al. |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2605.12038 |
| Download | |
| Code | https://github.com/showlab/OmniHumanoid |
Abstract
Cross-embodiment video generation aims to transfer motions across different humanoid embodiments, such as human-to-robot and robot-to-robot, enabling scalable data generation for embodied intelligence. A major challenge in this setting is that motion dynamics are partly transferable across embodiments, whereas appearance and morphology remain embodiment-specific. Existing approaches often entangle these factors, and many require paired data for every target embodiment, which limits scalability to new robots. We present OmniHumanoid, a framework that factorizes transferable motion learning and embodiment-specific adaptation. Our method learns a shared motion transfer model from motion-aligned paired videos spanning multiple embodiments, while adapting to a new embodiment using only unpaired videos through lightweight embodiment-specific adapters. To reduce interference between motion transfer and embodiment adaptation, we further introduce a branch-isolated attention design that separates motion conditioning from embodiment-specific modulation. In addition, we construct a synthetic cross-embodiment dataset with motion-aligned paired videos rendered across diverse humanoid assets, scenes, and viewpoints. Experiments on both synthetic and real-world benchmarks show that OmniHumanoid achieves strong motion fidelity and embodiment consistency, while enabling scalable adaptation to unseen humanoid embodiments without retraining the shared motion model.
Engineering Breakdown
Plain English
OmniHumanoid is a video generation framework that transfers motion between different robot/humanoid bodies without requiring paired training data for each new target embodiment. The key innovation is factorizing motion dynamics (which transfer across bodies) from appearance and morphology (which are body-specific), allowing the system to learn from multi-embodiment paired data once, then adapt to new robots using only unpaired video.
Key Engineering Insight
The critical engineering win is decoupling motion transfer from embodiment adaptation — you train a shared motion model once across multiple robots, then cheaply adapt to new embodiments without collecting paired data. This reduces the data collection burden from O(n²) robot pairs to roughly O(n) total embodiments.
Why It Matters for Engineers
Robot learning teams today spend enormous resources collecting paired human-robot or robot-robot demonstrations for each new morphology. This paper directly addresses that bottleneck by enabling synthetic data generation for new robots without expensive re-pairing, which accelerates sim-to-real pipelines and makes embodied AI training economically viable at scale.
Research Context
Prior cross-embodiment work either required paired data for every target robot (doesn't scale) or tried to learn everything end-to-end (entangles motion with morphology). OmniHumanoid advances the field by introducing a two-stage approach: shared motion learning from existing paired datasets, then lightweight unpaired adaptation. This enables streaming generation across arbitrary humanoid morphologies, a fundamental requirement for embodied AI platforms.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::