If you want to understand why humanoid robots still aren’t common in factories, watch what happens the moment a production line shifts from a fixed sequence to something that unfolds over dozens of steps. Most models can handle the immediate decision. They can look at a scene and plan a short move. But they’re not great at imagining what the scene will look like five or ten steps later, or what could go wrong in between. That lack of foresight makes long-horizon tasks — the kind real factories are full of — surprisingly brittle.
Enter Thinker-WM and the LIBERO benchmark
UBTECH’s answer is a newly released embodied world model called Thinker-WM. Built on top of the company’s existing foundation model, Thinker, it’s meant to function as a physical AI backbone, and it already has one concrete result to point to: first place on the LIBERO benchmark, a widely referenced test for embodied intelligence.
Long-horizon task execution is exactly where LIBERO separates systems that look promising from systems that actually work. Thinker-WM topped the leaderboard by solving two things at once. One is that environment states drift over time — lighting changes, objects shift, earlier actions leave traces that the robot has to account for. The other is error accumulation. Each step carries a small mistake, and over a long task those mistakes stack up. The world model counters both by projecting how a scene will evolve and adjusting the robot’s plan before execution drifts too far. UBTECH says this approach has produced the strongest known performance on these long sequences.
A unified architecture that plans and acts in the same pass
The architecture itself is a unified multimodal setup that optimizes visual perception and action generation together, rather than treating them as separate problems stitched together later. As the model imagines possible futures, it also refines how each action connects to the next, which is what makes long, multi-step operations look less jerky and more coherent.
Where the real edge comes from: a self-improving data loop
But the real story, the one that will matter to anyone trying to deploy these systems at scale, is how the model gets its data. UBTECH has put together a data collection network spread across multiple sites in China, capturing high-quality interaction data — industrial sorting, material transport, fine manipulation, two-arm coordination — all cleaned and filtered with enough care that the model can extract genuine physical patterns, not just surface correlations.
From there, Thinker-WM does something more interesting than simply memorise. Using a relatively small set of real-world reference data, the world model generates large volumes of high-fidelity synthetic training data. These aren’t random variations. They’re targeted: edge cases that real robots rarely encounter during normal data collection, dynamic disturbances, long-tail scenarios, tricky multi-step combos. The sort of stuff that would take forever to gather physically.
That synthetic data is then fed into downstream vision-language-action (VLA) models to improve fine motor control, responsiveness to unexpected changes, and closed-loop execution. The loop closes when those VLA models encounter failures or produce new interaction data during real tasks — that material flows back into the world model, sharpening its physical predictions and the quality of future synthetic scenes. One system gets better at imagining the world, the other gets better at acting in it, and neither needs a human to manually connect the dots each time.
The setup moves Thinker-WM past the stage of simply receiving data. It actively generates its own training material and refines its understanding of physics through a feedback loop with the real world. UBTECH says the model can keep improving its skills without constant manual intervention.
Opening the model to speed up real-world deployment
The company plans to open-source Thinker-WM soon through its developer community, Thinker-Cosmos, with an invitation to developers worldwide to help expand the data ecosystem. The stated goal: push humanoid robots into large-scale use across industries faster than would be possible if every lab worked in isolation.
