Dream Labs Raises $90M and Opens Public API for Dream-1 Robot Model

The seed round was jointly led by B Capital and Kindred Ventures, with participation from A.Capital, Unusual Ventures, Wing VC, SV Angel, Scribble VC and angel investor Adrian Aoun.

Dream Labs is emerging from stealth with Dream-1, which it describes as a “universal brain” for robots. The company says the model can control bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3 per hour.

The startup is targeting a lack of intelligence in current robots. It believes there are more than 5 million industrial robots working in factories and warehouses globally. Most are designed to perform a single task well but cannot be trained for other tasks or pass on what they learn to other machines, according to the company.

Robotic hardware costs have fallen, with highly dextrous arms selling for a few thousand dollars, but those products do not ship with intelligence. Buyers must teach an arm a specific job, a process that can take months of tedious, specialized programming; when new hardware arrives, the process has to be redone from scratch, Dream Labs says.

Dream-1 treats robotic skills as video prediction problems. It learns from what co-founder and CEO Joel Jang calls “human egocentric training data”—videos of people performing physical tasks such as stacking shelves and packing items into boxes. By studying the videos, Dream-1 learns to predict what should happen next and then determines the physical action the robot must take to make it happen. Dream Labs says video data is easier to accumulate than machine-specific datasets, allowing the model to transfer intelligence across different types of robotic hardware.

Jang said Dream-1 is designed to work with any two-armed robot equipped with parallel grippers; getting started requires a video feed and natural-language instructions. “The hardware is ready, but what’s been missing is a brain 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 past driving experience does the rest. Dream-1 approaches a new robot the same way, so it can do useful work in minutes, not months.”

The public Dream API lets developers translate visual inputs and natural-language prompts into robotic actions, according to the company. Access to the most powerful version of Dream-1 costs $3 per hour; a lightweight version costs 75 cents per hour. Through a Fine-Tuning-as-a-Service offering, developers can submit video demonstrating a new task—more footage is better—and Dream Labs says it will produce a customized version of Dream-1 within a few hours.

Dream Labs also launched a Robot Playground that anyone can access through a web browser; users can type a command and watch a live robot respond in real time, the company said.

Dream Labs is working with more than 20 robotic hardware companies whose machines cost between $2,000 and $10,000. The company says Dream-1 is designed to work with any robotic arm and that it works with partners to optimize the model.