Skild AI shows the Skild Brain learning skills from videos of humans

Skild AI/Skild Brain

Skild AI demonstrated its foundation model learning new tasks mostly from human videos, requiring less than an hour of robot-specific data. The approach escapes the robot-data bottleneck: teleoperation is slow to scale, while human instructional video is abundant.

• Model learns from internet-scale video (instructional, egocentric) rather than teleoperated demonstrations.
• Omni-bodied training across ~100,000 simulated robot morphologies enables transfer to unseen body types.
• Under 1 hour of robot data needed per task, versus 50–100 hours of teleoperation in prior approaches.
• Bridges embodiment gap: human movements map to different robot kinematics (arms, quadrupeds, mobile bases).