Dream-1 controls previously unseen bimanual robots through a public API

What is changing?

Dream Labs released its Dream-1 World Action Model through a public API on October 19. The company says the model can operate bimanual robots and perform tasks it did not see during training without demonstrations from that specific robot. Dream Labs also emerged from stealth with the announcement and disclosed a $90 million seed round.

Rather than being tied only to joint angles, Dream-1 can predict the position and open-or-closed state of the grippers, while each robot’s inverse-kinematics solver turns those targets into motion. The provider says it evaluated the model on 27 unseen robot embodiments. Results can vary with camera placement, workspace and inverse-kinematics quality.

API pricing is $3 per hour for Dream-1 and $2 per hour for the lighter Dream-1-flash. The model remains a research beta and has not been validated for safety-critical use. Where zero-shot transfer is insufficient, Fine-tuning-as-a-Service is intended to build a custom control model from a short recording of the robot’s own data.