Alibaba debuts RynnWorld-4D world model

By shifting world models from 2D pixel space to a structured 4D (RGB + depth + optical flow) representation, RynnWorld-4D solves the critical speed and alignment issues that have historically prevented diffusion-based world models from running closed-loop on real robots.

* Structured 4D vs. 2D: Adding depth and optical flow directly into the generative process provides the geometric structure and motion vectors required for precise robotic interaction, resolving the visual-spatial ambiguity of 2D world models.

* Low-Latency Action Generation: Generating control commands directly from the model's internal latents without waiting for multi-step visual denoising runs at 9Hz+, which is fast enough for reactive closed-loop manipulation.

* Large-scale Training: The use of 254.4M frames in Rynn4DDataset 1.0 represents a massive scale-up, helping the model generalize to diverse manipulation scenarios.

* Hardware Requirements: Generating multi-modal outputs (RGB-DF) and running diffusion models in real-time remains highly compute-intensive, potentially limiting deployment to edge devices without powerful local GPUs.