Dream Labs leaves stealth with $90 million and a model that controls robots it has never seen

Dream Labs came out of stealth on October 19, 2026 with $90 million in seed funding and Dream-1, a robot control model it offers through a public API. Founded about six months ago by former NVIDIA researchers and backed by B Capital and Kindred Ventures, the company says Dream-1 can operate two-armed robots with parallel grippers that it was never trained on, doing tasks it was never shown, with no fine-tuning or robot-specific data needed to start.

The trick is how the model sees the job. Dream Labs calls it a World Action Model: it learns to predict video and actions together, so every robot shows up simply as shapes moving in a camera view. Instead of commanding one robot's exact joints, Dream-1 predicts where each gripper should go, and the robot's own math turns that into motion, which lets the same model drive many different bodies. The company showed it controlling 27 unseen robots in simulation, including a research robot first released 18 years ago.

An independent evaluation by Robocurve placed Dream-1 above Opus 5.5 and Gemini 4 Argon on a robot control benchmark, at 16 times lower cost per attempt than Gemini 4 Argon. Running the model costs $3 an hour, or $2 for a lighter version. Dream Labs calls it a research beta, not validated for safety-critical use, and offers paid fine-tuning when zero-shot control falls short.

Why it matters

General-purpose robot software that works on hardware it has never seen could break the deadlock where every new robot needs its own training data, and put useful robots in reach of small teams.

Source: Dream Labs