Gasgoo Munich- UBTECH's homegrown unified multimodal world model, Thinker-WM, has claimed the top spot across three major embodied world model benchmarks. On the WorldOlympiad benchmark, the model posted a composite score of 0.740, leading all entrants in both 3D consistency and interaction fidelity. It also notched a total score of 4.66 on the EWMBench evaluation, taking the lead in core robotic-motion metrics.
On the WorldArena1.0 Track-1 simulation-environment benchmark, Thinker-WM swept all 12 key evaluation dimensions—including visual quality, content consistency, physical plausibility, 3D geometric precision, and instruction adherence. The model is built around a unified dynamic Transformer core, encoding RGB images, metric depth, and robotic actions into a single intermediate representation space for joint temporal modeling.
Image credit: UBTECH
This unified multimodal representation weaves vision, geometry, and actions into a shared latent space for deep coupling—rather than merely stitching them at the modality interface. Built on this architecture, the model can flexibly toggle between conditional and generative modalities, unlocking six capability modes: joint video-action generation, world prediction, policy reasoning, video-to-action translation, and inverse dynamics.
Fueling this unified representation is approximately 30,000 hours of heterogeneous training data—spanning 20,000 hours of first-person human video, 10,000 hours of real-robot data, and 1,000 hours of simulation data. In RoboTwin single-task ablation studies, adding the depth modality boosted the robot’s task-success rate by 15 percentage points over the depth-free baseline.
Lightweight Motion Probe experiments further confirmed that even without action supervision, the model’s visual representations already encode motion information tied to actions. Thinker-WM has now been deployed and validated on UBTECH’s robotic platforms. Going forward, the company plans to continuously update the model’s technical reports and subsequent iterations within the developer community, alongside rolling out an open plan for secondary development on the world model.
