Event

UBTECH Releases Unified Multimodal World Model Thinker-WM

1 reports · 1 sources

UBTECH's Thinker Team has released the unified multimodal world model Thinker-WM, which jointly models and generates RGB, metric depth, and robot actions through a shared Unified Dynamics Transformer; the same set of parameters supports video-action joint generation, world prediction, policy reasoning, action-conditioned generation, video-to-action generation, and inverse dynamics, among other functions. The model was trained on approximately 20,000 hours of first-person video, 10,000 hours of real robot data, and 1,000 hours of simulation data, of which approximately 4,000 hours of real robot data carry aligned depth supervision. The paper states that it ranks first on three benchmarks—EWMBench, WorldOlympiad, and WorldArena—and that a lightweight motion probe shows that frozen visual representations contain recoverable motion information.