Humanoid Robots Hit 100,000-Hour Data Bottleneck as 90% Sim Scores Fall to 12% in Reality
Updated
Updated · Yahoo Finance UK · Aug 22
Humanoid Robots Hit 100,000-Hour Data Bottleneck as 90% Sim Scores Fall to 12% in Reality
3 articles · Updated · Yahoo Finance UK · Aug 22
Summary
About 100,000 hours of robot manipulation data exist globally, far short of the scale that powered ChatGPT, making physical interaction the main bottleneck for humanoid learning.
Stanford tests showed robots that scored nearly 90% in simulation completed just 12% of real-world tasks, underscoring the persistent sim-to-real gap.
Current hardware adds another limit: no commercial humanoid can work an eight-hour shift on one charge, and leading models last only two to four hours; factory grippers also drop 5% to 15% of unfamiliar objects.
Unitree CEO Wang Xingxing still says a major software leap may take five to 10 years, while roughly $6 billion flowed into world-model companies in Q1 2026 to scale training through simulation rather than larger robot fleets.