Researchers have introduced DreamX-Phi 1.0, an action-conditioned video world model designed to predict future visual observations for robotic manipulation based on input frames, language instructions, and action sequences containing gripper states and end-effector poses. To prevent visual hallucinations where predicted movements deviate from commanded paths, the authors implement PRoPE-style geometric encoding, which injects per-arm SE(3) transformations into the attention mechanism to preserve rigid-motion structures and arm identity. To improve scene geometry and object persistence, the architecture incorporates a lightweight depth branch and utilizes SAM3 masks paired with a frozen V-JEPA teacher to maintain consistency during grasping tasks. For efficient deployment, the multi-step generator is compressed into a few-step student model through distribution-matching distillation. The model demonstrated high performance in competitive benchmarks, securing first place on Track 1 and second place on Track 2 of the WorldArena 2.0 Challenge.