UBTECH Thinker-WM World Model Tops Libero Benchmark, Solving Long-Horizon Robot Tasks
Release and Top Ranking: Addressing the Industry Core Weakness
As humanoid robot basic capabilities have improved, traditional industry solutions still face significant shortcomings in dynamic production line and other real-world environments. Most models can only perform static decision-making and short-horizon task decomposition based on real-time scenes, lacking the ability to simulate and predict future states of the physical world.
On May 5, 2026, UBTECH Technology announced its new embodied AI world model Thinker-WM, which achieved top ranking on the authoritative evaluation benchmark Libero with advantages in spatial generalization, long-horizon tasks, and comprehensive performance. Competitors on the same leaderboard included models from NVIDIA, Physical Intelligence, and Xiaomi.
Technical Core: Unified Architecture and Think Before Act Capability
Thinker-WM constructs a Diffusion Transformer unified multimodal spatial architecture. This technical path gives it significant advantages in complex task planning and execution:
- Multimodal Collaborative Optimization: During Flow Matching iterative evolution, the model achieves collaborative optimization of video representation and robot action space.
- Intelligent Imagination and Reasoning: Capable of intelligent imagination and reasoning for future scenes, progressively refining the rationality and coherence of action generation.
This Think Before Act capability is equivalent to equipping robots with a brain that can predict changes in the physical world.
Data Flywheel: Breaking Through Bottlenecks with Intelligent Generation
Training such models requires massive high-quality data. UBTECH builds an AI data flywheel system combining high-quality real data foundation with intelligently generated virtual data.
Performance Verification
In addition to winning the Libero simulation environment, UBTECH tested Thinker-WM using another authoritative benchmark Robotwin, achieving high-precision capabilities in action execution accuracy and trajectory stability.
Industry Implications
Thinker-WM topping the Libero benchmark surpassing NVIDIA and Physical Intelligence reflects intense technological competition in the embodied AI field.
