Dream Labs raises $90 million and opens an API for its cross-robot model Dream-1

Dream Labs, a San Francisco startup, has raised $90 million in seed funding. It has also launched Dream API, a public interface that takes a robot's camera feed and an instruction in plain English and returns motor commands. The company says this works even for robots its model has never seen.

Most robot intelligence today is built for one machine and one job. Dream Labs claims its flagship model, Dream-1, breaks that link. If the claim holds, the skill would sit in a shared model instead of inside each robot, and teaching a robot would look less like programming and more like asking.

The problem Dream Labs describes

According to Dream Labs, factories worldwide run a record 5 million industrial robots, and nearly every one is programmed for a single task. Most cannot pass what they have learned to other machines. The company says low-cost robots now sell for a few thousand dollars but ship with no built-in intelligence. Teaching one robot one task can take specialists months, it says, and a new task or new hardware often means starting from zero.

Daisy Cai, general partner at B Capital, put the shift in hardware costs this way: "A capable robot arm has gone from the price of a luxury car to the price of a high-end laptop. What it needs now is intelligence."

How Dream-1 is meant to work

Dream-1 is trained on video of people performing everyday tasks. Dream Labs says the model learns how the physical world works from that video. It predicts what will happen next and works out the action the robot should take. Because it learns from video rather than from data collected on a single machine, the company claims its capabilities transfer across hardware.

Dream Labs says Dream-1 can control any robot with grippers "out of the box", using zero-shot learning: no task-specific training data for the new machine. It needs no extra data to control a two-armed robot with parallel grippers, according to the company.

Co-founder and CEO Joel Jang compared it to driving a rental car. "When you get into a rental car, you don't relearn how to drive. You find the controls, and your past driving experience does the rest," he said. "Dream-1 approaches a new robot the same way, so a new robot can do useful work in minutes, not months."

The current scope is narrow. Dream-1 handles tabletop tasks such as picking and placing objects and folding laundry. Supported robots cost roughly $2,000 to $10,000. Support for wheeled mobile robots and dexterous hands is planned but not yet available.

What developers get

Dream Labs offers three products. Dream API lets developers sign up, connect a robot with grippers, and have it carry out a task described in natural language. The models run in Dream Labs' data centers around the world, which the company says keeps response times low wherever the robot is.

For unsupported hardware or new capabilities, the company sells fine-tuning as a service. Developers and robot makers submit a small batch of demonstration data and receive a tailored model within hours, it says. For its more than 20 hardware partners, Dream Labs refines the model for each robot.

The third product is a browser playground. Anyone can type commands to a robot in the company's San Francisco office and watch it respond live. Jang framed this as a deliberate choice: "Instead of edited demos, we're giving people a working model they can test themselves, mistakes included." That matters because robot learning is a field where polished demo videos often say little about reliability. A public, unedited test gives outsiders a way to judge the claims for themselves.

One hardware partner has given an early account. Ryan Chan, co-founder and CEO of MakerMods, which sells $2,000 bimanual robot arms, said: "With Dream-1, our robot was up and running in a few minutes and attempting tasks we never programmed. We can focus on building better hardware, and our customers get intelligence that keeps improving." Chan says "attempting", which is not the same as completing.

The funding

B Capital and Kindred Ventures led the round. A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and Adrian Aoun also took part, among others. Dream Labs says it will spend the money mainly on compute: to widen the range of tasks and robots Dream-1 can handle, and to support more users. According to Cai, the team did some of the defining work in robot learning at NVIDIA.

The company has not published success rates, benchmark results or pricing in the announcement. The claims about zero-shot control and transfer across hardware are Dream Labs' own and have not been independently checked.

What it means for work

The first job Dream-1 targets is the specialist work of teaching a robot a task, which the company says can take months. If a camera feed and a sentence are enough, that work moves from robotics engineers to anyone with an API key, and skill-building moves from each manufacturer to whoever runs the shared model. Hardware makers like MakerMods say they could focus on the machine and leave the behaviour to a model they rent. For now the reach is limited to tabletop picking, placing and folding on low-cost arms with grippers, which is far from the factory floor of 5 million single-task robots the company describes. What remains open is reliability: how often these robots finish a task correctly, and whether that is good enough for paid work, the announcement does not say.