Dream Labs

The product

What Dream Labs is building

Dream Labs is productizing the World Action Model, a class of robot model its team named at NVIDIA in February 2026 with the DreamZero paper. A WAM predicts video and actions together: it learns how the world moves from footage of people doing everyday tasks, then uses that prediction of what happens next to choose motor commands. Controlling the robot and imagining its future become one problem rather than two stacked models. The flagship is Dream-1. It takes three camera views (head plus a wrist camera on each arm), the robot's state and a plain-language task, and returns 1.6 seconds of actions per call, either as joint angles or as end-effector poses that each robot's own IK solver turns into joint commands. That shared end-effector space is the trick that lets one model drive many bodies. Dream Labs reports zero-shot control of 27 unseen bimanual parallel-gripper robots in simulation (PR2, ALOHA, Franka, KUKA iiwa, YuMi, UR5e and others) and real runs on arms like the OpenYAM, PiPER and Maker Arm. The business is unusually developer-shaped for robotics. The Dream API is pay-per-hour ($3/hour for Dream-1, $2 for Dream-1-flash). Fine-tuning as a service adapts the model to a new robot with as little as five minutes of that robot's data in their examples. A browser Robot Playground drives a real arm in the SF office. More than 20 hardware partners sell supported robots in the roughly $2,000 to $10,000 range. They are candid about limits: Dream-1 is a "research beta" that "will make mistakes," is not validated for safety-critical use, and today covers tabletop work like pick-and-place and laundry folding. Wheeled mobile manipulators and dexterous hands are next.

The money

Why this matters

B Capital's Daisy Cai put the thesis in one line: "A capable robot arm has gone from the price of a luxury car to the price of a high-end laptop." Once hardware is a $2K-$10K commodity, the scarce thing is a policy that works on any of it without a robotics PhD and a month of teleoperation per task. That is the gap Joel Jang names directly: "The hardware is ready. What's been missing is a brain any robot can use." The other shift is the data source. Robot-collected demonstrations are slow and expensive; human video is effectively unlimited. A model that turns human video into transferable action priors, and then needs minutes rather than days of robot data to adapt, changes who can build a robot product. A $90M seed is a compute bet that this scaling curve holds, and an API-first launch means real usage data from day one.

Investors: B Capital (Daisy Cai, GP) and Kindred Ventures (co-leads), A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, Angels: Adrian Aoun