Dream-1 controls robots it was never trained on

Dream Labs came out of stealth with $90 million in seed funding and Dream-1, a model it calls a World Action Model, offered through a public API as a research beta. The claim that stands out is zero-shot control of robot arms the model never saw during training. The team reports testing 27 unseen bimanual, parallel-gripper robots in simulation and 4 real ones, using no embodiment-specific data for those runs. The model predicts video and actions together from three cameras, a head view plus a wrist view on each arm, and outputs end-effector poses that each robot's own inverse kinematics solver turns into joint commands. That design is also its main limit, because performance drops on robots whose kinematics differ sharply from the training set.

An independent evaluation by Robocurve on October 18 placed Dream-1 ahead of Opus 5.5 and Gemini 4 Argon on a Tier 1 benchmark, citing 16 times lower cost per task attempt than Gemini 4 Argon and 1.7 times faster responses than GPT-6 Astra. Pricing is $3 per hour for Dream-1 and $2 for a faster variant, and each call returns 1.6 seconds of actions. Fine-tuning on a robot's own data helped further, with five minutes of data enough for some robots while others needed up to 98 minutes.

Why it matters

If you work in robotics, a model that drives hardware it was never trained on, through a plain API, lets you prototype without first collecting a dataset for each robot. Read the kinematics caveat and the research-beta label closely before you rely on it for anything safety-critical.