San Francisco-based Dream Labs has raised $90 million in seed funding to develop AI models capable of controlling robots from different manufacturers without requiring extensive training for each machine.
The round was led by B Capital, alongside Kindred Ventures, with participation from A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and angel investor Adrian Aoun.
Founded in 2026 by former Nvidia researcher Joel Jang, the company emerged from work on DreamZero, a robotics model introduced in February 2026. Jang previously worked on Nvidia's Project GR00T, an initiative developing foundation models for humanoid robots.
Dream Labs is addressing one of robotics' biggest limitations: most robots are trained or programmed for particular tasks, environments and hardware configurations. Even when a robot learns a new skill, transferring that capability to another machine typically requires additional data collection and training.
Its first commercial offering, Dream-1, uses a World Action Model architecture that jointly predicts visual information and physical actions. Instead of relying exclusively on robot-specific motion data, the model learns representations of how objects and environments change during physical interactions.
Alongside the funding, Dream Labs launched Dream API, a public interface that accepts visual observations and natural-language instructions and returns actions a robot can execute. The company is positioning itself as an intelligence provider for robotics manufacturers rather than building its own fleet of robots.
Dream Labs reports working with more than 20 hardware partners. Its research demonstrations include zero-shot control across different two-armed robots with parallel grippers, including evaluations on 27 previously unseen robot configurations in simulation.
The company also offers fine-tuning services for hardware and tasks where the general model performs inadequately. In demonstrations, it showed improvements after collecting as little as five minutes of additional robot-specific data, although more complex configurations required considerably more.
Dream-1 is accessible through an API starting at $3 per hour of model operation. That pricing covers access to the model, not the robot hardware or the full cost of operating an autonomous system.
The company expects to direct much of its new capital toward computing infrastructure, model training, additional hardware support and broader task coverage.
The wider opportunity is significant. As robot hardware becomes cheaper, general-purpose intelligence could become a separate commercial layer, allowing manufacturers to focus on mechanical design while buying software capabilities from specialized AI companies.
However, Dream-1 remains a research beta, and the company explicitly states that it has not been validated for safety-critical applications. Its current capabilities also remain concentrated on compatible robotic configurations rather than arbitrary machines.
The central test is whether Dream Labs can move from impressive demonstrations to repeatable deployments with economically viable success rates. A general-purpose robotics API becomes commercially valuable only when developers can rely on its performance across changing environments without extensive customization.