LiberAI launches Liber-0, pretrained on 2.2 million hours

LiberAI, founded in December 2025, is positioning Liber-0 within the emerging generation of foundation models for robotics, though it specifically calls its system a World Action Model. The company's central bet is that human experience, which exists on an enormous scale compared with robot-collected demonstrations, can serve as the primary training resource for general-purpose robot manipulation.

The technical challenge is bridging what LiberAI describes as three gaps: a visual gap between human recordings and robot cameras, an action-supervision gap where hand positions inferred from ordinary video can be incomplete, and an embodiment gap between human and robot bodies. Its response combines large-scale human video, robot demonstrations, and instrumented-glove recordings that capture hand poses and trajectories alongside first-person and wrist-mounted camera views. Because glove-based collection does not require a robot at every location, LiberAI sees it as a potentially more scalable source of physical AI training data.

LiberAI has also designed its LEO robot with dexterous hands, articulated wrists, and an adjustable waist to reduce unnecessary differences between human and robot movement. The company says it is exploring tactile and force sensing as future dimensions for its data gloves, noting that contact and force can produce very different results even when two movements look identical. Its longer-term objective is for accumulated human experience to produce increasingly general physical capabilities, reducing the robot-specific demonstration data needed to teach new behaviors.