Hugging Face Releases RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions

Summary:Hugging Face has released RynnWorld-Latent, a cross-embodiment video world model that leverages the 3D-grounded latent actions from RynnLAM as a shared conditioning interface. This approach overcomes the challenge of incompatible action spaces between robot and human videos. The model is pretrained on 23 datasets totaling approximately 553.5 million frames and fine-tuned for target robots through post-training. Experiments demonstrate improved video prediction metrics on public benchmarks and en

---

Technical Mechanism Analysis

RynnWorld-Latent is built on a pretrained video transformer architecture, combining global action tokens with an identity-initialized local pathway to align actions with future video latents. The key innovations include:

1. Cross-Embodiment Conditioning: Leveraging the 3D-grounded latent actions from RynnLAM as a shared interface to address the incompatibility of action spaces between robot and human videos.
2. Data Efficiency: Pretraining on 23 datasets totaling approximately 553.5 million frames, encompassing human interaction videos, robot demonstrations, and simulation data.
3. Post-Training Adaptation: Adapting the model to target robots through post-training, supporting two data-generation routes: digital teleoperation and latent-action synthesis.

Engineering Trade-offs and Performance

• Strengths:

  • Achieves efficient utilization of heterogeneous data, reducing reliance on specific action space data.
  • Demonstrates significant improvements in video prediction metrics on public benchmarks.
  • Shows excellent performance on real robots, enabling precise action-following capabilities.

• Challenges and Trade-offs:

  • High demand for high-quality annotated data, leading to increased training costs.
  • Post-training adaptation requires customization for different robots, adding to deployment complexity.

Developer Implementation and Deployment Recommendations

1. Data Preparation: Developers are advised to prepare diverse datasets, including human interaction videos, robot demonstrations, and simulation data, to fully leverage the model's cross-embodiment capabilities.
2. Post-Training Adjustment: Conduct post-training adjustments for target robots to ensure optimal performance in specific application scenarios.
3. Performance Optimization: In resource-constrained environments, consider model compression or quantization techniques to reduce computational costs.

Conclusion

RynnWorld-Latent demonstrates the potential of cross-embodiment world models in handling heterogeneous data, providing new technical solutions for robot control and simulation. Its strong performance in video prediction and action-following makes it a promising tool for various robotic applications.

---

Source: Hugging Face Daily Papers (2026-10-27)

— END —

Tags: #Hugging Face Follow #Cross-Embodiment Model Follow #Video World Model Follow #Robot Control Follow #Latent Actions Follow