Dream Labs, from the NVIDIA DreamZero team, raises $90M and opens an API for robots

Dream Labs, a startup founded by the DreamZero team, has emerged from stealth with two announcements at once. The first is a $90M seed round, a very large start for a company with no public sales history. The second is the launch of Dream API, which its creators call the world's first public API for a robot "brain." Ge Yunhao (葛云皓) is the chief technology officer. The idea looks simple: you connect the API to your robot, write a task in plain language, and the system decides what to do.

What exactly launched

Until now, access to models that control physical machines has looked different. Either these are closed internal systems at large labs, or open-source checkpoints that you have to deploy yourself, fine-tune for specific hardware, and maintain. Dream Labs offers a scheme that is familiar to developers: a request goes to the API, and the response comes back as actions for the robot.

Three facts follow from the available part of the announcement. Access is public, not limited to a closed list of partners. The control interface is built on natural language. The API can be connected to your own robot, which means the company does not sell a finished machine. It sells a layer of intelligence.

Where the technology comes from

DreamZero is known as an NVIDIA research project built around so-called world action models. The core of the approach is that the model does not simply match an image to a command. It relies on an internal representation of how the world will change after an action. The approach borrows ideas from video generation models. The practical goal is to transfer skills to new tasks and environments without retraining for each one.

If Dream API really builds on this legacy, the bet is clear. The main problem in robotics today is not a shortage of hardware but brittle software. A robot that folds towels perfectly in the lab often gets lost in someone else's kitchen. A universal "brain" behind an API could remove that brittleness, or at least move it to the cloud, where it can be fixed centrally.

What is known from the announcement

• Round: seed, $90M
• Product: Dream API, public access
• Input: an instruction in natural language
• Output: control of the connected robot
• CTO: Ge Yunhao (葛云皓)
• Origin: the DreamZero team
• Not disclosed in the available material: investors, pricing, the list of supported platforms, response latency, task success metrics

The last item matters more than the others. For a robot, latency is critical: if a command goes to the cloud and back, fast reactions, such as holding on to an object, will have to stay on board. So the architecture most likely splits the work between cloud planning and a local controller. But that is an inference, not a stated fact.

How developers should approach it

It is too early to rewrite your stack around Dream API, but it is worth testing. A practical sequence looks like this:

• Pick one narrow task, such as moving objects or simple sorting, and phrase it with three or four different instructions. This shows how robust the system is to wording.
• Measure the latency from command to first movement and compare it with the requirements of your mechanics. If it exceeds a few hundred milliseconds, you need a local safety layer.
• Ask the company which manipulators and mobile platforms are supported and what integration requires: calibration, cameras, control format.
• Plan for fault tolerance in advance. A cloud service controls a physical device, so if the connection drops, the robot must stop safely.
• Ask for the pricing terms. Whether you pay per request or per minute of operation will determine if the API suits mass-produced products or only prototypes.

The main question for Dream Labs is simple: how well does the "brain" work on a robot it has never seen? Independent tests will answer that, not a presentation. If the answer is convincing, hardware startups will have a way to avoid hiring a separate team for training control policies. If not, the $90M will become another expensive attempt to turn a research demo into infrastructure.