Dream Labs, a San Francisco start-up founded by former NVIDIA robotics researchers, has opened Dream-1, a robot foundation model, to developers through a public API. Its investor Wing VC says the company came out of stealth in October with $90m in seed funding.
Dream-1 is what the company calls a world action model: it predicts video and robot movements together. It takes a camera feed and a plain-English instruction and returns about 1.6 seconds of arm movements at a time. For now it is built for two-armed robots with simple two-finger grippers.
The headline claim is that Dream-1 can control robots it was never trained on. In the company’s own tests, that covered 27 unseen robots in simulation and four real ones. Dream Labs also says about five minutes of demonstration data improved results on two robots, though robots that differed more needed 77 to 98 minutes.
The company is open about the limits. Performance varies by robot, the model does not control how an arm reaches a position, and Dream Labs calls it a research beta that “has not been validated for safety-critical use”. Pricing starts at $3 an hour.
Why it matters: most robot brains come bundled with a company’s own hardware. Dream Labs is betting developers will rent one over the internet instead.
Sources: Dream Labs, Dream-1 research page; Wing VC investment note