UBTECH has published details of Thinker-WM, a world model developed by its Thinker team that encodes RGB video, metric depth and robot actions in a single shared representation. A technical paper is available on arXiv. The company says the model ranks first on three embodied world model benchmarks. Those rankings are UBTECH's own claims, based on results available when it made the announcement.
By UBTECH's account, Thinker-WM scored 0.740 overall on WorldOlympiad, a benchmark built on real robot interaction data that covers more than ten robot bodies and several camera setups. It also posted the top single scores for 3D consistency (0.531) and interaction fidelity (0.765). On EWMBench it reported a total of 4.66, ahead of models including Qwen-RobotWorld and Sora 2, with leading results on trajectory-matching and dynamics metrics. On WorldArena 1.0 Track-1, a simulation benchmark, the company says it led across 12 evaluation dimensions.
The design argument is that many world models are stitched together: video generation handles pixels, a policy network handles actions and a separate module predicts geometry. UBTECH contends that this splits signals that are physically coupled. Thinker-WM instead maps each modality through its own encoder into a common latent space and models them jointly in a shared Unified Dynamics Transformer before decoding video, depth maps and executable actions. Depth is treated as a native output rather than an auxiliary signal. In single-task ablations on RoboTwin, adding depth raised manipulation success by 15 percentage points over a baseline without it.
Because the modalities share one set of weights, swapping which inputs are given and which are generated lets the same model run six modes, including joint video-action generation, future prediction, policy inference, video-to-action and inverse dynamics.
Training data totals roughly 30,000 hours: about 20,000 hours of first-person human video from homes, retail and factories, about 10,000 hours of real robot data from several robot types, and about 1,000 hours of simulation. All robot actions are expressed as end-effector poses, with a body-specific tag so different machines share one generation space. A data pipeline called Data Factory handles sampling, layered quality checks and depth-video alignment.
To test whether the shared space truly couples modalities, the team trained a small motion probe on frozen visual features to recover end-effector displacement, which UBTECH presents as evidence of cross-modal coupling.
SEE ALSO: WorldArena 2.0 at IROS 2026 · UBTECH and FAW-Volkswagen Take Humanoid Robots Into Factory Logistics