Dream Labs' first World Action Model, available through a public API. By predicting video and actions together, Dream-1 controls robots it has never seen, on tasks it was never shown, with no data from that robot.

A 5B video-generation model trained to predict 3D scene flow and robot actions. Rendering scene flow as RGB lets the pretrained video VAE encode it directly, and predicting this motion rather than future video is far more sample-efficient for learning actions.

A framework for quantifying the transfer value of a robotic manipulation task to other tasks in a taxonomy. Shows that both foundational tasks and complex tasks accelerate learning on others, but only tasks higher in the hierarchy can be used to improve performance on simpler tasks.