Curate what matters
Shot detection, motion-aware segmentation, technical quality checks, and multimodal review retain complete physical events.
Video world model for Physical AI
StrucPhysVideo
moves like the world.
StrucPhysVideo generates high-fidelity video while learning the regularities beneath it—persistence, contact, motion, and state change.
object persistence
continuous motion
45.5%Physics-IQ Verified
#1 / 17models evaluated
TI2V + IA2Vlanguage and action conditioned
01 — Generated worlds
See physical intelligence
See physical intelligence
unfold over time.
The complete set of 21 selected generations across physical phenomena, everyday scenes, robot manipulation, and action-conditioned prediction.
02 — How it works
From observable interaction
From observable interaction
to predictive world model.
StrucPhysVideo connects interaction-rich data, grounded descriptions, and action conditioning in a single Physical AI program.
Describe what changes
Structured labels ground objects, actions, camera motion, and temporally ordered state transitions in visible evidence.
Predict what follows
Text-image and end-effector action conditions guide generation while preserving visual quality and temporal continuity.
TI2VText + image → video
IA2VImage + action → video
03 — Quantitative results
Leading physical
Leading physical
prediction.
45.5%
Physics-IQ Verified Score
+2.8 pts over the next-ranked model
Physics-IQ Verified
Snapshot · 16 Sep 2026Image-to-video leaderboard
* indicates our reproduction. Scores for the other comparison models are taken from the Physics-IQ benchmark snapshot dated 16 Sep 2026. Verified Score combines spatial overlap, spatiotemporal overlap, weighted spatial overlap, and normalized pixel error. Higher is better.
Grounded supervision compounds.
33.2%Wan2.2-5B+8.4 pts vs. reproduced baseline
45.5%LingBot-Video-30B+7.6 pts vs. reproduced baseline
Higher frame fidelity, action conditioned.
PSNR ↑23.89
SSIM ↑0.8721
LPIPS ↓0.0880
DROID · 253-frame rollout
Long-horizon action following
Relative scores normalized to the bidirectional baseline; higher is uniformly better.
Trajectory3.88×
Depth3.52×
Interaction1.83×
LPIPS⁻¹3.30×
Evaluation note The LingBot-Video result is reproduced by our team; the other comparison-model scores reflect the Physics-IQ Verified benchmark snapshot dated 16 Sep 2026. AgiBot results are controlled training-distribution diagnostics under each model’s native conditioning interface; they do not establish held-out generalization or isolate architecture alone.
StrucPhysVideoGenerate the world.
Back to top ↑