DreamX-Phi: Separate Robot Video Prediction from Execution

Reconstructed and revised on 2026-09-30 from the surviving Chinese manuscript and checked sources. This is not a verbatim recovery of the previous English article.

What the model predicts

The DreamX-Phi paper predicts future observations from an observed frame, language instruction and an action sequence containing end-effector poses and gripper states. It uses per-arm geometric encoding and object-consistency constraints. This is not verified general-purpose video editing or evidence of physical execution. VQOS did not run the model or a robot.

Label industrial footage clearly

Distinguish filmed operation, simulation, generated illustration and concepts. The product owner should verify the advertised capability. A generated successful grasp cannot serve as equipment test evidence. Disclosures must be visible to the audience, not confined to internal notes.

Check the arm and object

Inspect the acting arm, gripper state, contact, position and action order. Smooth motion may still use the wrong arm or lose the object. Without genuine action inputs, do not claim compliance with a robot control trajectory.

Obtain source test footage and task descriptions from the product owner rather than filling evidence gaps with generation.

Deliver facts with versions

Record each segment's source, purpose and approver. Recheck affected scenes when model numbers, motion or copy changes. Separate illustrative footage from technical validation materials in the delivery list.

Read VQOS services and specify source evidence and labeling in your brief. Research rankings do not replace equipment acceptance, execution logs or safety review.