Recently, UBTECH ROBOTICS' self-developed unified multimodal world model Thinker-WM claimed first place across all three widely recognized embodied world model benchmark evaluations.
On the WorldOlympiad benchmark, Thinker-WM ranked first with a composite score of 0.740, while its 3D consistency score of 0.531 and interaction fidelity score of 0.765 each placed first among all participating models. On the EWMBench benchmark, Thinker-WM achieved an overall score of 4.66, surpassing mainstream models such as Qwen-RobotWorld and Sora2. In the simulation environment benchmark WorldArena1.0 Track-1, Thinker-WM led overall across 12 key evaluation dimensions, including visual quality, content consistency, physical plausibility, 3D geometric accuracy, and instruction following.
According to the company, unlike the "stitched-together" approaches commonly used by previous world models, Thinker-WM is built around a Unified Dynamic Transformer (UDT) with a unified representation layer. It brings RGB images, metric depth, and robot actions into a single shared dynamic representation for joint evolution, integrating multiple capabilities into one system, including joint video-action generation, world prediction, policy reasoning, video-to-action conversion, and inverse dynamics. In a RoboTwin single-task ablation experiment, the manipulation success rate rose by 15 percentage points after introducing the depth modality compared with a baseline without depth.
On the training data side, Thinker-WM's corpus includes approximately 20,000 hours of first-person human video, about 10,000 hours of real robot data, and roughly 1,000 hours of simulation data. The model uniformly uses end-effector poses to represent robot actions and embeds dedicated labels for different robots.
UBTECH ROBOTICS stated that Thinker-WM has completed deployment validation on the UBTECH ROBOTICS robot platform, and the related technology will continue to empower next-generation embodied intelligence systems.
Comments