Skip to main content
DreamZero jointly predicts future video and robot actions from camera observations, measured state, and a language task. Reactor runs inference and maintains the episode’s temporal context; your client captures observations and controls the robot.

​
Choose an embodiment

The variants share a streaming lifecycle, but their checkpoints, camera names, joint order, and state commands differ. Select an embodiment before adapting a client. Matching an action width alone does not establish compatibility with another robot. Both variants emit action_chunk messages once a prompt and all camera streams are available. They do not wait for an explicit predict command or execution acknowledgement. Your application decides how to replace the pending plan, which validated targets to execute, and when to replan. Optional predicted video is diagnostic output, not a sensed observation or a safety check. The upstream DreamZero project supplies the model family. The YAM endpoint uses Robocurve’s MolmoAct2 BimanualYAM fine-tune; it is a separate checkpoint from the DROID endpoint.