DreamX-Phi 1.0 Wins Track 1 of the WorldArena 2.0 Challenge

There is a compact paper from the DreamX Team on arXiv that is worth sitting with, not because it is loud but because of what it is doing structurally. DreamX-Phi 1.0 is an action-conditioned video world model for robotic manipulation: you give it an observed frame, a language instruction, and a prescribed action sequence of end-effector poses and gripper states, and it predicts the future observations that would follow.

The interesting design choice is how they get geometry into the model. Rather than treat the arm as another patch of pixels to be predicted, they inject per-arm SE(3) transformations into attention via what they call PRoPE-style geometric encoding, meant to preserve arm identity and rigid-motion structure. A lightweight depth branch handles scene-level geometry. To keep the thing being manipulated from dissolving mid-grasp, they pair SAM3 masks with a frozen V-JEPA teacher, using it to maintain object consistency throughout grasping. Read together, those choices are an argument that manipulation video is not generic video and should not be modeled as such.

The headline claim in the abstract is that at the time of writing, DreamX-Phi 1.0 achieves first place on Track 1 and second place on Track 2 of the WorldArena 2.0 Challenge. That is the load-bearing external validation in the paper.

The abstract does not publish numbers beyond those two rankings, so the size of the Track 1 lead, the identity of the Track 2 winner, and the specific robots and tasks WorldArena 2.0 uses are not visible from what has been posted. A world-model paper also does not, by itself, show that better next-frame prediction translates into better downstream policies on a real robot, and this one stops at the leaderboard.

If the code and models drop as the team says they will, the more useful contribution may not be the ranking but the recipe: a demonstration that pushing arm kinematics into the attention pattern, rather than hoping a generic video transformer discovers them, is a live path for manipulation-specific world models.