UBTECH unveils Thinker-WM, a world model unifying video, depth and robot actions

UBTECH detailed Thinker-WM, a world model that encodes RGB video, metric depth and robot actions in a single shared latent space. The company claims top results on three embodied world model benchmarks: 0.740 on WorldOlympiad, 4.66 on EWMBench and leadership across 12 dimensions of WorldArena 1.0. Training used roughly 30,000 hours of data, including 20,000 hours of human video and 10,000 hours of real robot data.

• 0.740 on WorldOlympiad and 4.66 on EWMBench, ahead of Qwen-RobotWorld and Sora 2
• Adding depth raised manipulation success by 15 percentage points in RoboTwin tests
• Trained on ~30,000 hours: 20,000 human video, 10,000 robot, 1,000 simulation
• One model runs six modes, from video-action generation to inverse dynamics