←Home

    Robots That Learn by Watching People

    How embodied AI turns recordings of human hands and first-person experience into robot skills: capturing the data, recovering hand and contact signals, extracting latent actions and physical knowledge, and bridging the gap between human and robot bodies.

    Background: Vision-language-action models · World models

    In order

    1 stories

    Photo: Yang Plasticine

    Embodied · Egocentric datasets

    Two thousand hours of smartphone video of human hands, packaged for training robots

    Open-AoE gathers phone-recorded footage of people handling objects, annotates hand poses and action steps, and supplies tools to turn it into robot training data.

    arXiv preprint · 15 July 2026