Midcentury passes 2M hours of egocentric human data, banks $15M seed for robot training

CyberRobooo · x · 2026-09-27

NYC physical AI startup Midcentury claims its first-person human behavioral dataset has passed 2 million hours, covering 50+ environments and 20,000+ tasks, with hands visible in 90%+ of footage. It includes 3D hand tracking, depth maps, task annotations, plus 50K hours of gameplay and 69K hours of multilingual dialogue. A cloud simulation platform, Matrix, lets humanoids train directly on this data. Citing Dyna Robotics' results that enough egocentric data lets AI grasp physical-world regularities, the milestone helped Midcentury raise a $15M seed — human behavioral data is becoming the "new oil" of physical AI.

Related event: Midcentury Hits 2 Million Hours of Egocentric Human Data for Robotics(2 posts)→

Original post →

More from Venture

Venture channel →