UBTech's in-house unified multimodal world model, Thinker-WM, has taken first place across all three of the field's recognized benchmarks for embodied world models: a composite score of 0.740 on WorldOlympiad, a total score of 4.66 on EWMBench, and an overall lead across 12 key dimensions in Track-1 of WorldArena1.0.

Three Report Cards: Not Just First Overall, But First in Every Category

WorldOlympiad is built on a large body of real-world robot interaction data, spanning more than ten robot bodies and multiple camera configurations, with an emphasis on testing generalization across morphologies and viewpoints. Thinker-WM topped the benchmark with an overall score of 0.740, including 0.531 on 3D consistency and 0.765 on interaction fidelity — the best result among all models tested on both individual measures.

EWMBench evaluates three dimensions: scene consistency, the correctness of robot end-effector motion, and semantic alignment with instructions. Thinker-WM scored 4.66 overall, ahead of mainstream models including Qwen-RobotWorld and Sora2. It also led on the core robot-motion metrics, including nDTW (normalized dynamic time warping), which reflects how well trajectories match in space and time, and Dyn (dynamics consistency), which captures whether movement speed and acceleration are plausible.

WorldArena1.0 Track-1 is a benchmark for simulated environments, and Thinker-WM leads overall across its 12 evaluation dimensions, including visual quality, content consistency, physical plausibility, 3D geometric accuracy, and instruction following.

The problem with the old approach: things that are physically coupled got split across three separate pipelines

Getting robots to do real work isn't about recognizing what's in front of them. It's about anticipating what happens next. That's the problem world models are built to solve.

Until now, though, world models have mostly been stitched together: a video generation module handles the visuals, a policy network handles the actions, a geometry predictor handles spatial information, each computing on its own before the outputs are joined through interfaces. The trouble is that actions reshape the environment, the environment constrains actions, and geometric structure feeds into both. Signals that are tightly coupled in the physical world get split across separate modules, and the causal links between them disappear.

UBTech offers an analogy: a cup being pushed off a table. In language, it's a sentence. In motion, it's a string of joint commands. In video, it's movement across a frame. In depth, it's a continuously shifting field of distances. Four representations point to the same event, and generating any one of them amounts to interrogating the same underlying process.

But the four representations overlap only partially. Action carries neither appearance nor context: the same joint commands mean something entirely different on an empty table than on a cluttered one. Language carries no spatiotemporal detail — "the cup was pushed off the table" says nothing about which side it was pushed from, how much force was used, the exact path of the hand, or how many times the cup tumbled. Video carries no explicit geometry or metrics: it tells you neither how big the cup is nor how far away. Depth carries no appearance, semantics, or intent; it tells you where surfaces sit, how they occlude one another, how they make contact, but not whether the thing is a cup or a tablecloth, and certainly not why anyone would push it.

A Different Route: Not Merging at the Interface, but Unifying at the Representation Layer

Thinker-WM's approach runs RGB video, metric depth and robot actions through separate encoders that map each into a shared latent space. Those representations feed into a common Unified Dynamic Transformer (UDT) for joint temporal modeling, after which dedicated decoders reconstruct the video, the depth maps and executable robot actions.

That means multimodal interaction happens inside the Transformer itself: visual tokens, geometry tokens and action tokens exchange information directly through the same attention mechanism, rather than being stitched together at the input or output layer. Appearance, geometry and action are no longer three separate pipelines, but three expressions of a single process of world evolution.

Depth is no longer "auxiliary supervision" but a native generative modality

One pivotal choice in this architecture: geometric information is no longer treated merely as an auxiliary supervision signal. Thinker-WM makes depth a native generative modality, sharing a consistent latent topology and spatiotemporal organization with video, so depth supervision shapes the shared representation directly. Physical spatial structure becomes the world model's native language rather than a supplementary module bolted on after the fact.

The change shows up in measurable robot performance: in a single-task ablation on RoboTwin, adding the depth modality lifted manipulation success rates by 15 percentage points over a baseline without depth. In other words, spatial geometry does not just help the model make sense of a scene; it directly improves action generation and task execution.

Six Capabilities, One Set of Weights

Because all modalities share a single set of parameters, Thinker-WM can support multiple capability modes on the same weights simply by switching which modality serves as the condition and which is generated: it can anticipate and imagine future frames, work backwards to infer a robot's execution trajectory, and extract action information from video. The modes listed include joint video-action generation, world prediction, policy reasoning, video-to-action, and inverse dynamics.

The foundation rests on 30,000 hours of heterogeneous data.

Thinker-WM's training corpus is substantial by the standards of embodied AI:

  • Roughly 20,000 hours of first-person human video spanning household, commercial retail, and industrial settings, filling in the diversity of tasks, objects, and scenes that teleoperation struggles to capture cheaply, paired with about 1,000 hours of aligned depth data;
  • Roughly 10,000 hours of real-world robot data: drawn from a range of robots with different form factors, it pairs visual observations with executable actions as supervision, including about 3,000 hours with aligned depth annotations;
  • Roughly 1,000 hours of simulation data: providing high-precision geometric and physical supervision signals.

Action spaces, coordinate systems and control protocols vary enormously from robot to robot. Thinker-WM extends its unifying logic to the data layer: all robot actions are expressed as end-effector poses, while each robot gets its own dedicated tag embedded in the tokens. Machines of different morphologies can therefore share a single generation space without losing the precise embodied control each one needs.

Turning that data into actual model capability relies on a data infrastructure layer called Data Factory, which decouples data management from training-time sampling and handles three jobs. It organizes video, state, action and language information under a single structure, generating the samples each training mode requires on demand. It applies tiered quality checks to filter out anomalous clips and verify that depth data aligns with the video. And it stages data in advance through asynchronous prefetching and reuse of encodings, cutting idle waiting and redundant computation.

Motion Probe: Proving It Really Learns Coupling in a Shared Space

"Unification at the representation layer" sounds like a claim about architecture, but one question is hard to avoid: is the model genuinely learning how vision, geometry, and action couple together inside a shared space, or is it simply processing several streams of information in parallel and placing them side by side?

UBTech designed a lightweight Motion Probe experiment to answer a single question: attach a tiny motion-prediction head to frozen UDT visual representations, train only that head, and see whether end-effector displacement can be recovered from purely visual features. The probe's prediction error came out low — even without action supervision, the visual representations already encode motion information tied to action. That points to a direct conclusion: visual perception and motion planning are not two separate pipelines but are deeply coupled within a shared latent space.

What this means

Thinker-WM's significance goes beyond topping benchmarks. It points toward a new paradigm for embodied world models: a world model should not be a crude pairing of "video generator plus policy head," but a single unified generative process. Visual observation, spatial structure, and action planning are, after all, just different expressions of the same physical-world process.

When a robot can reason simultaneously about changing visuals, spatial structure, and motion trajectories within the same representational space, its grasp of the physical world's causal logic takes a step forward. That is also the critical leap that moves robots from merely executing tasks to actually understanding the real world.

According to UBTech, Thinker-WM has completed deployment validation on the company's robot platform, and the technology will continue to feed into its next-generation embodied intelligence systems.