RynnWorld-4D: A 4D Embodied World Model for Robot Manipulation

Haoyu Zhao · hf · 2026-07-08

The paper proposes a multimodal 4D world model that simultaneously generates RGB videos, depth maps, and optical flow data from a single RGB-D image and language instruction. It achieves efficient robot manipulation by integrating a unified diffusion process with inverse dynamics policy learning. By unifying multiple sensing modalities within a single generative framework, this model lowers the training data barrier for robotic manipulation policies.

Related event: RynnWorld-4D: Embodied World Model for Robot Manipulation(3 posts)→

Original post →

More from Embodied

Embodied channel →