Dream Labs raises $90 million to use simple video feeds to help build smarter robots

Dream Labs, a startup that aims to make robots smarter and capable of performing a wide variety of tasks, has raised $90 million in seed funding, coinciding with the launch of a public application programming interface that provides access to its basic universal robot model.

Today’s round was co-led by B Capital and Kindred Ventures and included participation from other companies including A.Capital, Unusual Ventures, Wing VC, SV Angel, Scribble VC and angel investor Adrian Aoun.

Dream Labs targets the lack of intelligence in today’s robots, which are quite numerous but also quite limited in many environments. The startup assumes that more than five million industrial robots work in factories and warehouses worldwide. These robots certainly do useful work, otherwise their owners wouldn’t have invested in them, but Dream Labs co-founder and CEO Joel Jang believes they should be able to do much more.

What stops them is a lack of expertise and knowledge. Almost all of these five million robots are designed to perform a single task, and they do it very well. However, they cannot be trained for other tasks, nor can they pass on the knowledge they have acquired to train other robots and machines.

While the cost of robotic hardware has fallen dramatically and highly dexterous robotic arms now sell for just a few thousand dollars, these products lack any intelligence. Rather, it is the buyer’s responsibility to teach this robotic arm how to perform a specific task, and that is a process that can take many months as it requires a lot of tedious and highly specialized programming work. Worse, whenever new robotic hardware is introduced, the entire process has to be done from scratch.

Today, Dream Labs is breaking cover with Dream-1, a robotic foundation model that controls bimanual robots it’s never seen before, on tasks it’s never been shown before, in places it’s never been shown before, starting at $3 an hour.

Dream API is available to everyone today. 🧵 pic.twitter.com/MB1snhUOAJ

— Dream Labs (@dreamlabs_ai) October 19, 2026

Jang believes there should be an easier way to train robots to perform new tasks, and his solution is a highly versatile base model called Dream-1, designed to be a “universal brain” for robots. Dream-1 treats robotic skills as simple video prediction problems. It teaches robots what Jang calls “human egocentric training data.” It’s a fancy term for videos of people performing physical tasks like stacking shelves and packing items into boxes.

By studying these videos, Dream-1 learns to predict what should logically happen next and then, based on that, determines the physical action the robot needs to take to ensure that it actually happens. The big advantage of teaching robots in this way is that video-based data is much easier to collect than specialized machine-specific data sets. This means Dream-1 can quickly collect the data it needs to teach robots new skills and seamlessly transfer information to different types of robot hardware.

Jang said Dream-1 is designed to work with any two-armed robot equipped with parallel grippers. All it requires to get started is a video feed and natural language instructions – basically someone telling it what to do. “The hardware is ready, but what’s missing is a brain that any robot can use,” Jang said. “When you get into a rental car, you don’t relearn how to drive. You find the controls and your previous driving experience does the rest. Dream-1 is approaching a new robot in the same way, so it can do useful work in minutes, not months.”

The next step is to make this intelligence widely available to developers, and that’s where today’s API launch comes into play. Through the Dream API, developers can access Dream-1 and immediately begin translating visual and natural language inputs into robot actions. Dream Labs charges developers $3 per hour to access the most powerful version of Dream-1, while prices drop to just 75 cents per hour for a lightweight version.

With Dream-1, robots can also be taught completely new skills that the model was not previously aware of through the startup’s fine-tuning-as-a-service offering. Developers simply need to submit a video demonstrating a new task (the more video, the better), and Dream Labs will use it to create a highly customized version of Dream-1 within a few hours.

To demonstrate the capabilities of its model, the company also launched the Robot Playground, which anyone can access via a web browser. Simply enter a command and watch a live robot react in real time.

Dream Labs currently works with more than 20 robotics hardware companies, with machines costing between $2,000 and $10,000. While Dream-1 is designed to work with any robotic arm, the company works with these partners to optimize the model for their specific hardware. This work allows for a better overview of what each robot can do, enabling better performance.

Jang said Dream-1 is already very capable at tasks like picking up and placing objects and folding clothes, but it has much bigger ambitions in the long term. For example, he plans to teach the model to support mobile robots that use wheels, as well as robots with fully dexterous or human-like hands. Money from today’s round will help fund the computing power needed to make that happen, he said.

B Capital’s Daisy Cai said robotic arms can now be purchased for the same price as a high-end laptop. “What they need now is intelligence,” she continued. “This team has done some of the critical work in robot learning at Nvidia, and now anyone with an API key can build on it.”

Photos: Dream Labs