Want to make your dumb robot smart? $3 an hour is all it costs. Or 75 cents, if you’re okay with a slightly less capable physical AI model. Dream Labs, a San Francisco startup founded this year by former leaders of Nvidia’s GR00T robotics research team, announced a $90 million seed round today. Dream Labs also launched the Dream API, which it says is the first publicly available API for a robot foundation model.

It’s kind of magic: send the API your robot’s camera feed plus some plain-English instructions, and it essentially operates your robot … even if it has never seen that particular model before. The team behind the startup says this is the missing link in physical AI.

“The hardware is ready. What’s been missing is a brain any robot can use,” CEO and co-founder Joel Jang said in a statement. “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. Dream-1 approaches a new robot the same way, so a new robot can do useful work in minutes, not months.”

The model behind it is called Dream-1. Dream-1 is trained on video of people doing everyday tasks (just like Figure with Index or Neura Robotics in Germany). The model predicts what will happen next in a scene, then works out what the robot should do about it. Because it learns from human video rather than from one specific machine with one specific set of instruction software, Dream Labs says the skills transfer across hardware.

And they’ve got a big claim: out of the box, Dream-1 can control any robot with two arms and parallel grippers. Essentially instantly, or zero-shot, as roboticists say.

Cross-embodiment transfer – teaching robot B to do what robot A is already doing – is one of the hardest problems in robotics. Almost every robot on the planet is specifically trained for its specific body and its specific task. What one machine learns rarely transfers to another, which has slightly different arms, hands, motors, geometry, friction, and more. There are about 5 million industrial robots working in factories globally, according to Dream Labs, and nearly all of them are programmed for a single task.

Six months ago, Jang says, adapting the team’s model to a new robot took 30 minutes of data from that robot. Before that, getting anything working could take 100,000 hours.

“That was six months ago. Now we just need the robot.”

The trick is treating robot control as a video problem. “If the AI model can predict a video of a robot doing it, then it can do it,” Jang says. “So now control becomes a video prediction problem.”

So how well does it work?

Dream Labs says zero-shot pick-and-place succeeds about 80% of the time, with other tasks landing in the 50% to 60% range. That’s not good enough for a factory line, but Dream Labs isn’t aiming at the factory line, at least not yet. Many robot model companies are chasing industrial applications, Jang told me, “getting to high nines, you know, 99% success rates of specialized tasks.” Instead, he wants “everyday capabilities for everyday robots.”

To achieve production-ready quality, the company offers fine-tuning as a service. Developers or robot makers submit demonstration data; they get back a tailored model within hours.

That’s a big deal: robot training in hours, not weeks, or months. And for a few dollars an hour, not five or six-figure integration consultant fees. Today, Jang says, a developer who buys a $5,000 or $10,000 robot has to grab open-source models like GR00T or π, reserve GPUs, train, and host the model. “So it takes like three to six months just to solve a single problem.” Dream Labs will also refine the model for each of its partners (the company currently has “more than 20” hardware partners).

Ultimately, this enables heterogeneous fleets: one model running robots from several different makers. That’s super-useful: companies can buy the robots they need, not settle for other models from the same manufacturer due to hoped-for integration ease, or not worry that running a mixed fleet will double or triple training and updating costs and time.

Dream Labs says the exact mix of training data that it has used is proprietary for now, but that it’s using a lot of human egocentric data: video shot from a person’s point of view. That’s the same bet others are making. Nature Will said last week that its Natura-1 hand model was pre-trained on over a million hours of it, and Figure is spending a billion dollars on a gig platform paying people to film themselves doing chores.

Software and training are clearly getting cheaper with this API. The hardware side of the equation is getting cheaper too. Robots supported by this Dream Labs release cost between $2,000 and $10,000 today, and dropping.

“A capable robot arm has gone from the price of a luxury car to the price of a high-end laptop,” says Daisy Cai, a general partner at B Capital, one of the investors in the round. “What it needs now is intelligence.”

B Capital led the $90 million seed round, and A.Capital, Kindred Ventures, Unusual Ventures, SV Angel, Wing VC, Scribble VC and Adrian Aoun also participated.

The Dream API also launched today will cost $3 an hour for the flagship model and $0.75 an hour for the smaller flash model. This follows an established model in AI now: frontier or best models are expensive; mid-grade models with lower levels of capability are cheaper. Developers and integrators can pick the model they need depending on task complexity.

This is an interesting inversion of how we usually think about robot costs: ongoing versus upfront.

Run a robot around the clock on the flagship and you’ll pay just over $26,000 a year. Use the cheaper model and it’ll run you roughly $6,600. That adds some significant yearly cost to your $2,000 to $10,000 robot, but a smart robot should be able to do better and more efficient work than a brainless one. Most robots won’t run nonstop, further reducing potential costs, and if the AI wars have taught us anything, it’s that token prices always come down over time thanks to efficiencies and competition.

Dream Labs’ competition is stiff and well-funded.

Skild AI has raised well over a billion dollars for its robot-agnostic Skild Brain, Physical Intelligence has its π models, Google has Gemini Robotics, and Nvidia, of course, has GR00T. Some in the industry, like General Robotics CEO Ashish Kapoor, think this middle layer of robot foundation models is going to get squeezed between frontier models and job-specific post-training.

Dream Labs’ answer to that might be distribution.

Sign up, get an API key, connect a robot, pay by the hour. That’s how developers adopted large language models, and it’s a sharp contrast to robot brains that are only available inside one company’s hardware, or an open source model that you install on your own hardware.

There’s also something new, cool, and radically transparent. Dream Labs’ Robot Playground lets anyone with a browser type commands to a real robot in the company’s San Francisco office and watch it respond, live and unedited. It’s try-before-you-buy, or try before you rent, and it should give potential customers a good, quick, and free glimpse at what the models could do for them.

“Instead of edited demos, we’re giving people a working model they can test themselves, mistakes included,” Jang said.

To start, Dream-1 handles tabletop tasks with grippers, like picking and placing objects and folding laundry. Dexterous hands work with the model too, Jang says, but cost is the problem: “they’re $100,000 just for a pair.” He expects capable hands in the $5,000 to $10,000 range by next year, and Dream Labs will add more support as they arrive.

Support for wheeled mobile robots and dexterous hands is coming next, and most of the $90 million will go to compute to expand the tasks and robots Dream-1 can handle.