Dream Labs raises a $90M seed and opens an API to its universal robot foundation model

Dream Labs has closed a $90 million seed round, an unusually large first raise even by 2026 standards, co-led by B Capital and Kindred Ventures. Alongside the funding, the company opened a public API to what it calls a universal robot foundation model. The core claim is that the model can learn robotic behavior from ordinary video feeds instead of relying on expensive teleoperation or motion-capture datasets.

The technical bet is the same one the large labs are making: video is the most abundant source of physical-world data, and a model that maps video demonstrations to actions sidesteps robotics' data bottleneck. Opening an API at the seed stage signals that Dream Labs wants developers and hardware makers building on its model early. That is a platform play rather than a vertically integrated robot company.

The competitive field is crowded and well funded. Google DeepMind's Gemini 4 Argon, launched last week, converts video demonstrations and sketches into MuJoCo physics code and robot motor actions, and lifted Boston Dynamics Spot manipulation success from 56% to 75% in trials. A standalone startup has to show it can match or specialize beyond frontier labs that fold robotics into general models. Its pitch will likely rest on hardware neutrality and focus.

Skepticism is warranted. Community threads this week argued that general physical robotics is still premature for messy real-world work such as home repairs, and raised liability questions about who pays when an AI-directed action causes damage. Video-learned policies also tend to struggle with fine contact dynamics that video does not capture. Watch for published success rates on real hardware and for which robot makers sign on as API customers.