UBTECH Thinker-WM: Unified Multimodal World Model
Overview
This research presents Thinker-WM, a unified world model that jointly models RGB, metric depth, and robot action through a shared intermediate representation. The model supports multiple functions including video-action generation, world prediction, and policy inference, and achieves top performance on several embodied intelligence benchmarks.
From arXiv Robotics — research abstracts
Event evidence and corrections
1 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · training_hours_human_video: 20000 hours · Basis not reported
“Thinker-WM 训练语料在具身智能领域具备可观体量： 约2万小时第一人称人类视频 ：覆盖家庭、商业零售、工业等场景，补足机器人遥操作难以低成本获取的任务、物体与场景多样性，同时配套约 1 千小时的对齐深度数据。”
Exact source · revision 2qbitai
Reported quantity · training_hours_robot_data: 10000 hours · Basis not reported
“约1万小时真实机器人数据 ：来自多款不同形态机器人，提供视觉观测与可执行动作的配对监督，其中约 3 千小时配套对齐深度标注。”
Exact source · revision 2qbitai
Reported quantity · world_olympiad_score: 0.74 other · Basis not reported
“< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400"> 2026-10-16 16:00:31 来源： 量子位 优必选 Thinker-WM在三大公认具身世界模型基准评测中全部拿下第一名 01 三大基准全榜登顶，Thinker-WM 斩获三项第一 日前，优必选 Thinker-WM在三大公认具身世界模型基准评测中全部拿下第一名，刷新行业成绩： WorldOlympiad ： Thinker‑WM 取得综合得分 0.740 位列榜首。”
Exact source · revision 2qbitai
Reported quantity · ewmbench_total_score: 4.66 other · Basis not reported
“EWMBench： Thinker‑WM 取得综合总分 4.66，超过 Qwen‑RobotWorld、Sora2 等主流模型。”
Exact source · revision 2qbitai
Reported quantity · three_d_consistency: 0.531 other · Basis not reported
“模型取得 3D 一致性 0.531、交互保真度 0.765，两项单项指标位居全部参评模型首位，体现模型对三维几何结构与机器人交互过程的精准建模能力。”
Exact source · revision 2qbitai
Reported quantity · interaction_fidelity: 0.765 other · Basis not reported
“模型取得 3D 一致性 0.531、交互保真度 0.765，两项单项指标位居全部参评模型首位，体现模型对三维几何结构与机器人交互过程的精准建模能力。”
Exact source · revision 2qbitai
Reported quantity · Reported duration: 20000 hours · Basis not reported · Differing source assertions
“duce a lightweight motion probe that shows the frozen visual representation contains recoverable motion information. Thinker-WM is trained on a heterogeneous corpus comprising approximately 20K hours of egocentric video, 10K hours of real-robot data, and 1K hours of simulation data, including approximately 4K hours of real-robot data with aligned depth supervision. Extensive experiments demonstrate strong performance”
Exact source · revision 4Source owner not reported
Reported quantity · Reported duration: 10000 hours · Basis not reported · Differing source assertions
“duce a lightweight motion probe that shows the frozen visual representation contains recoverable motion information. Thinker-WM is trained on a heterogeneous corpus comprising approximately 20K hours of egocentric video, 10K hours of real-robot data, and 1K hours of simulation data, including approximately 4K hours of real-robot data with aligned depth supervision. Extensive experiments demonstrate strong performance”
Exact source · revision 4Source owner not reported
Reported quantity · Reported duration: 1000 hours · Basis not reported · Differing source assertions
“duce a lightweight motion probe that shows the frozen visual representation contains recoverable motion information. Thinker-WM is trained on a heterogeneous corpus comprising approximately 20K hours of egocentric video, 10K hours of real-robot data, and 1K hours of simulation data, including approximately 4K hours of real-robot data with aligned depth supervision. Extensive experiments demonstrate strong performance”
Exact source · revision 4Source owner not reported
Reported quantity · Reported duration: 4000 hours · Basis not reported · Differing source assertions
“duce a lightweight motion probe that shows the frozen visual representation contains recoverable motion information. Thinker-WM is trained on a heterogeneous corpus comprising approximately 20K hours of egocentric video, 10K hours of real-robot data, and 1K hours of simulation data, including approximately 4K hours of real-robot data with aligned depth supervision. Extensive experiments demonstrate strong performance”
Exact source · revision 4Source owner not reported
Evidence dependency changes (5)
- 2026-10-30T23:04:44.178Z · evidence updated · source revision 4. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T23:04:25.694Z · evidence updated · source revision 4. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:44:43.309Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:44:36.859Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-21T20:08:06.337Z · evidence updated · source revision 4. Evidence extraction was updated; current source attributions are shown above.
Developments
- 2026-10-16 08:00 UTC · 1 reportsUBTECH Thinker-WM: Unified Multimodal World ModelQbitAI — Robotics：Dominating Three Benchmarks. UBTECH Thinker-WM: Unified Multi-modal Representation, Single Model Achieves Six Capabilities
- 2026-10-14 04:00 UTC · 1 reportsThinker-WM: A Unified World Model with Shared Multimodal Representations for EmbarXiv Robotics — research abstracts：Thinker-WM: A Unified World Model with Shared Multimodal Representations for Embodied Intelligence
Report timeline
Follow attributed reports and material updates.
- QbitAI — RoboticsSignalDominating Three Benchmarks. UBTECH Thinker-WM: Unified Multi-modal Representation, Single Model Achieves Six Capabilities
UBTECH Thinker-WM achieved first place in three major benchmark evaluations: WorldOlympiad, EWMBench, and WorldArena1.0 Track-1, demonstrating its advantage in cross-robotic form and perspective generalization capabilities.
- arXiv Robotics — research abstractsSignalThinker-WM: A Unified World Model with Shared Multimodal Representations for Embodied Intelligence
This research presents Thinker-WM, a unified world model that jointly models RGB, metric depth, and robot action through a shared intermediate representation. The model supports multiple functions including video-action generation, world prediction, and policy inference, and achieves top performance on several embodied intelligence benchmarks.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).