Dominating three major benchmarks! UBTECH Thinker-WM: Unified multimodal representation, a single model achieving six major capabilities

What happened: UBTECH's Thinker-WM, a unified multimodal world model, took first place on three embodied world model benchmarks: WorldOlympiad, EWMBench, and WorldArena1.0 Track-1. Key numbers: WorldOlympiad overall score 0.740, with 3D consistency 0.531 and interaction fidelity 0.765. EWMBench total 4.66. Why it matters: The model encodes RGB video, depth, and robot actions into one shared representation rather than stitching separate modules, and UBTECH says it has completed deployment validation on its robot platform.

UBTECH's Thinker-WM, a unified multimodal world model, took first place on three embodied world model benchmarks: WorldOlympiad, EWMBench, and WorldArena1.0 Track-1.

The model encodes RGB video, depth, and robot actions into one shared representation rather than stitching separate modules, and UBTECH says it has completed deployment validation on its robot platform.