Physix Frontier · News Briefing Card (TMTPost · Oct 14, 2026)

Teams Explore Pretraining Robot Manipulation on Human Video

KEY FACTS

On October 10, Xingdong Era released the weights of VPP2, a two-stage video model covering task prediction and visual planning.
Dyna-2 used a key-turning experiment to compare pretraining scale; the 1-million-hour group succeeded 9 out of 10 tests.
After adding 12 minutes of human video, NVIDIA DreamZero raised average progress across 9 tasks from 38.3% to 54.3%.
τ0-WM used about 27,300 hours of heterogeneous data, of which 3,000 hours were first-person human video.
In RoboWM-Bench evaluation, after Wan 2.6 generated video and converted it to actions, drawer-closing success was 50% and picking success was 20%.

KEY DATA

9/10Dyna-2 key-turning success rate

38.3%→54.3%DreamZero task progress improvement

3000/27300 hoursτ0-WM human video share

50%RoboWM-Bench drawer-closing success rate

PHYSIX OBSERVATION

Human video is upgrading from auxiliary material to the main fuel for robot task understanding, while robot data retreats to body control. Dyna-2's key experiment and DreamZero's progress gain prove that scale works, but RoboWM-Bench's low success rates also expose the gap between visual plausibility and physical executability. The industry's next competitive point is who can first build a stable bridge from video prediction to joint actions.

Source: TMTPost report