Dream Labs Raises $90 Million to Let Robots Share Skills With Each Other
Dream Labs, a startup started by former NVIDIA robotics researchers, has raised $90 million in seed money. The company also launched a public tool, called an API, for its new model, Dream-1. The goal is simple: let one robot's learned skills work on a different robot too.
The funding round was led by B Capital and Kindred Ventures. A.Capital, Unusual Ventures, and other investors also took part.
The Problem Dream Labs Wants to Fix
Right now, if one robot learns a task, that knowledge does not transfer to a different robot. Each machine often has to learn things on its own, which costs time and money. Dream Labs wants developers to start with one shared model, then only add extra training when a specific robot or job truly needs it.
An API is a tool that lets different computer programs talk to each other. Here, it lets outside developers connect their robots to Dream Labs' model without building their own system from scratch.
Along with the API, Dream Labs launched a fine-tuning service (a way to further train the model on your own robot) and something called Robot Playground. This is a website where anyone can send commands to real robots sitting in the company's San Francisco office. Dream Labs says most of the new funding will go toward computing power as it grows the model's abilities.
How the Model Actually Works
Dream Labs calls Dream-1 a "World Action Model." It links two things together: guessing how a scene will change next, and deciding what the robot should do about it.
This matters because robot arms are not all built the same. A command that moves one robot's joints may not work on another's. So instead of controlling joints directly, Dream-1 predicts where the robot's gripper (the claw-like part that grabs things) should end up, and whether it should be open or closed. Each robot then uses its own internal system to turn that target into actual joint movements.
The model uses three camera views: one of the overall scene and one on each wrist. This lets it understand different-looking robot arms through sight, rather than needing identical hardware everywhere.
What the Testing Showed
Dream Labs tested the model on 27 robots it had never trained on before, inside a computer simulation, running 300 attempts for each one. Results were compared to YAM, the robot Dream-1 was originally trained on. Performance changed a lot depending on the robot, so the full detailed results matter more than the headline claim alone.
A transferred skill working somewhat is not the same as it working perfectly every time. Dream Labs itself separates "fully completed" tasks from "partial progress" in its results.
For now, this skill-sharing works best on two-armed robots with simple, parallel grippers. Developers who need more can use the fine-tuning service, which lets them upload their own robot's data and get a specially adapted model instead of starting over. Dream Labs shows this working with as little as five minutes of data in some cases, though other setups need much more.
Using the model costs money too. Dream Labs lists $3 per hour for Dream-1 and $2 per hour for a lighter, faster version called Dream-1-flash. Each use of the model returns 1.6 seconds of robot action.
What This Means for You
You do not need to understand robotics or AI jargon to see why this matters. If this works well, building and training robots for new jobs could get faster and cheaper, which could show up in manufacturing, delivery, warehouses, and more. Dream Labs says plainly that the model is still a research product and has not been checked for safety-critical use.
- If you run a business that uses or is considering robots, know that shared "skill" models like this are an early trend, not yet a proven standard.
- If you are simply curious, you can try the Robot Playground online to watch real robots respond to commands, without owning any equipment yourself.
- If you work in tech or investing, watch how much extra training real customers still need after adopting this model. That answer will shape whether this approach truly saves time and money.
Next Step
If you want to see this idea in action without any technical background, visit Dream Labs' Robot Playground and watch how a real robot responds to a simple instruction.
Source: Unite.AI
