A robotics startup called Dream Labs came out of stealth on Friday with $90 million in seed funding and the public launch of Dream-1, a model the company says can control robots it has never seen before, on tasks it was never shown, starting at $3 per hour through a public API.
The announcement, published on Dream Labs' research page, quickly reached the front page of Hacker News, where commenters debated whether zero-shot robot control is finally practical or still a research demo. The company describes Dream-1 as its first "World Action Model" available to anyone, marking a bid to make general-purpose robot behavior as pluggable as text and image generation became under large language models. For more context on this story, see our ongoing AI news.
What Dream-1 Actually Does
Give Dream-1 a bimanual robot with parallel grippers, connect it to the API, and the company says the model starts performing tasks out of the box — no fine-tuning, no demonstrations from that specific robot, and no machine learning expertise required.
"We want generalizable robotics to be accessible to anyone with a robot, not just large labs," the company writes in its launch post, describing the setup as close to plug-and-play: connect your robot, point it at the API, and run. The team says it runs weekly internal hackathons to keep the time from signup to a working robot as short as possible.
The company's launch material shows beta testers running Dream-1 across a range of tabletop tasks: sorting candy into containers at Anvil Robotics, pulling tissues at BitRobot in Shenzhen, stacking cups at Cortex in Singapore, folding boxes at KAIST in Seoul, and moving cubes between containers at Hugging Face's lab in Paris. Dream Labs says beta testers at 12 external sites ran the model autonomously at eight times normal speed, with in-house runs at four times speed.
A second set of demonstrations shows longer autonomous sessions: folding a dress, clearing dishes into a sink, putting eggs into a carton, restocking cans from a box, and wiping a countertop with a sponge.
From a 30-Minute Paper to a Company
Dream Labs traces its origin to February 2026, when its team released DreamZero, the research paper that introduced World Action Models — a paradigm the company says the robotics community has since adopted as one of the most promising approaches in robot learning. The result that surprised the team, according to the launch post, was that after post-training DreamZero on just 30 minutes of data from a YAM robot, the model generalized zero-shot to tasks it had never seen. That kind of transfer previously required hundreds or even thousands of hours of robot-specific training data.
After six months of building, the team says that result has matured into "early signs of zero-shot embodiment generalization," the capability Dream-1 now exposes through its API.
Why the Data Problem Matters
Robotics has long been stuck in what the company calls a chicken-and-egg loop: a robot needs to be useful before people buy it, but making it useful requires training data from that exact robot, which only accumulates once enough units are in the field. Self-driving cars escaped a version of this problem because human drivers generated data from day one. Most robots never had that luxury.
If a single model can generalize across robot bodies, a new machine starts with useful skills instead of an empty dataset — and every deployment adds training data that improves the next robot, the argument goes.
Cheap Hardware Is the Other Half of the Story
Dream Labs pairs its model claims with a hardware argument: capable robots are getting dramatically cheaper. A two-armed setup now costs around $2,000, the company says, and seven semi-humanoid robots are available for under $15,000 — a price tier that barely existed a year ago.
At the consumer end, Dream Labs points to Hugging Face's $399 Microduck desktop robot, which the company says sold more than 10,000 units within four days of its launch in late August. The distinction the company draws is deliberate: enterprise buyers make what it calls a spreadsheet decision, and a robot that cannot beat a worker on cost, speed, or reliability does not get purchased. Consumers are different — more than 10,000 people paid $399 in four days for a machine whose main job is to be delightful, one that walks, waddles, and roller-skates. Somewhere between that toy and a fully capable home robot, Dream Labs argues, lies a sweet spot where adoption starts to compound, and every household robot becomes a new source of training data.
The Caveats
All performance claims come from the company itself, and independent verification is limited to what beta partners have shown publicly. Dream Labs devotes its own section to what Dream-1 cannot do, an acknowledgment that generalization remains narrow: the demonstrated tasks are largely tabletop manipulation with gripper-equipped arms, a long way from the unstructured chaos of homes and construction sites.
Robotics history is full of striking demos that stalled outside the lab. But the combination of a public API, per-hour pricing, and a hardware price floor dropping toward consumer territory makes Dream-1 one of the more concrete attempts yet to turn robot foundation models into infrastructure other developers can simply call.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
