UBTECH says its Thinker-WM world model takes first place on three embodied AI benchmarks

UBTECH's unified world model claims the top spot on WorldOlympiad, EWMBench and WorldArena 1.0 - robots that predict, not just react.

UBTECH says its self-developed unified multimodal world model, Thinker-WM, has taken first place on three embodied world-model benchmarks. The company frames the result as a step toward robots that can predict what happens next and plan around it, rather than simply reacting to what their cameras see.

On the WorldOlympiad benchmark, the model scored 0.740 overall and led all entrants on 3D consistency and interaction fidelity. On EWMBench it reached a composite score of 4.66, topping the core robot-motion metrics. And on the WorldArena 1.0 Track-1 simulation-environment benchmark, it led across all 12 key evaluation dimensions, including visual quality, content consistency, physical plausibility, 3D geometric precision and instruction adherence.

The architecture centers on a unified dynamic Transformer that encodes RGB images, metric depth and robot actions into a single intermediate representation for joint temporal modeling, rather than stitching modalities together at the interface. That shared latent space lets the model switch between conditional and generative modes, unlocking capabilities that include joint video-action generation, world prediction, policy reasoning, video-to-action translation and inverse dynamics.

The representation rests on roughly 30,000 hours of heterogeneous training data: about 20,000 hours of first-person human video, 10,000 hours of real-robot data and 1,000 hours of simulation. In single-task ablation studies on RoboTwin, adding the depth modality lifted the robot's task-success rate by 15 percentage points over the depth-free baseline. UBTECH says Thinker-WM has completed deployment validation on its robot platforms, and it plans to publish the model's technical reports along with an open plan for secondary development on the world model.