Alibaba's DAMO Academy released an open-weights robot world model, RynnWorld-4D, on July 8 that generates a robot's predicted future as a simultaneous stream of color, depth, and motion, providing geometric and kinematic information before a physical action. The model addresses a limitation in current Vision-Language-Action (VLA) models like Google DeepMind's RT-2, Physical Intelligence's π₀, and NVIDIA's GR00T N1, which cannot simulate outcomes before moving and struggle with generalization beyond their training distribution. RynnWorld-4D generates three streams of information simultaneously: RGB color frames, depth maps, and optical flow fields, a combination called RGB-DF. The model takes a single RGB-D image and a natural-language instruction, then generates all three future-state streams inside a unified Video Diffusion Transformer, using the Wan2.2-TI2V-5B architecture with three parallel branches tied together through cross-modal attention layers and frame-wise three-dimensional Rotary Position Embeddings (3D RoPE).