en.Wedoany.com Reported - The field of embodied intelligence has seen significant progress. UBTECH, based on its self-developed large model Thinker, has launched a new embodied intelligence world model, Thinker-WM, aimed at enhancing the long-horizon task planning and execution capabilities of humanoid robots in real industrial scenarios. The model achieved the top position on the authoritative embodied intelligence evaluation benchmark Libero, competing alongside companies such as NVIDIA, Physical Intelligence, and Xiaomi.
Thinker-WM adopts a Diffusion Transformer unified multimodal architecture, achieving collaborative optimization of video representation and action space through Flow Matching, thereby possessing the ability to reason about future scenarios. Leveraging multiple data collection centers across various locations, UBTECH has accumulated real interaction data from industrial sorting, material handling, and other tasks, providing the model with a cognitive foundation of the physical world.
In the Libero evaluation, Thinker-WM particularly excels at long-horizon task execution, solving challenges such as temporal changes in environmental states and error accumulation. Furthermore, the model demonstrated high-precision action generation and video fidelity in the Robotwin benchmark test.
UBTECH has also built an AI data flywheel system, using a small amount of real data to generate a large volume of high-fidelity virtual data, reducing training costs and improving generalization performance. This model forms a bidirectional evolutionary closed loop with the VLA model, continuously optimizing physical reasoning accuracy. In the near future, UBTECH plans to open-source Thinker-WM in the developer community to promote the construction of the embodied intelligence ecosystem.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
