GR00T 2 is an open world-action model for robots that imagines short future videos and generates the matching motor actions. It handles chained multi-step tasks over minutes and adapts from human wearable or simulation data to real robots without task-specific teleoperation.