Ever wished your robot could just get it, no complex programming needed? Well, a new startup called Dream Labs, founded by ex-Nvidia robotics gurus, just secured a whopping $90 million to make that dream a reality. They’ve launched the Dream API, offering what they call the first publicly available API for a robot foundation model. Essentially, for as little as 75 cents an hour, you can send your robot’s camera feed and some plain-English instructions, and their AI will practically run the show – even if it’s never seen that specific robot before! It’s touted as the missing puzzle piece for truly smart, adaptable physical AI.
The secret sauce is their Dream-1 model, which learns from watching videos of humans doing everyday tasks. Instead of being tied to one specific machine, it predicts what should happen next and tells the robot how to act. This is a game-changer because, traditionally, robots are super specialized; what one learns rarely translates to another with slightly different parts. Dream Labs is tackling this “cross-embodiment transfer” challenge head-on, meaning a robot can be useful in minutes, not months, without relearning everything from scratch. While still in its early stages for some tasks, seeing 80% success for simple pick-and-place operations is pretty impressive!
And for those who need factory-level precision, Dream Labs offers a clever fine-tuning service. Just send them some demo data, and they’ll whip up a custom model for your robot in mere hours – a huge leap from the weeks or months it usually takes. This isn’t just about speed; it’s about cost, slashing the hefty integration fees typically associated with robotics. Ultimately, this paves the way for “heterogeneous fleets,” where a single AI brain can control robots from different manufacturers, letting companies pick the best bot for the job, rather than being stuck with one brand. They’ve already got over 20 hardware partners on board, so things are moving fast.