Dream Labs Raises $90M Seed for Robot Foundation Model API

The browser window is part of Dream Labs' pitch to the robotics market. Through its Robot Playground, someone without a machine can give instructions to a real robot in the company's San Francisco office and watch the attempt. A developer considering the technology gets a chance to bring a difficult request into the demonstration.

That invitation accompanies a $90M Seed financing announced on October 19, 2026, led by B Capital along with Kindred Ventures. The San Francisco startup is offering Dream-1, a robot foundation model, through a public API. Most of the new capital is intended for compute as Dream Labs expands the tasks, hardware and users it can support.

For robotics builders, the arrangement puts an outside model supplier into a consequential relationship: the one between a hardware maker and the person trying to put its machine to work. Dream Labs is asking that builder to buy intelligence as an evolving service, with adaptation available when the general model needs help on a particular robot or task.

What the Dream Labs seed round funds

Dream Labs' company-origin announcement names A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and Adrian Aoun among the other investors. B Capital's portfolio also lists the company. The funding gives Dream Labs resources for a compute-heavy service whose obligations continue after the initial model release, as developers begin sending it requests from their own equipment.

Founded in 2026, Dream Labs is led by co-founder and CEO Joel Jang, a former NVIDIA GR00T robotics researcher. Current reporting from 36kr identifies fellow co-founders Yunhao Ge as CTO, Scott Reed as chief scientist and You Liang Tan as chief robot officer. The company's team account traces its work to DreamZero, the World Action Model research released in February 2026.

The research history matters because the founders have chosen a commercial delivery model alongside a technical approach. An API gives other developers a route into the model without requiring them to operate the entire research stack themselves. It also gives Dream Labs a continuing responsibility for the infrastructure behind that access, making compute a practical part of the financing story.

That responsibility extends into staffing. Dream Labs' careers page lists openings across robot learning, ML systems, backend software, robotics and security in San Francisco. Those disciplines cover much of the distance between publishing an impressive result and keeping an external developer's service usable, although the company has not disclosed a numeric hiring target.

How Dream-1 connects intelligence to a robot

Dream-1 receives observations and instructions and produces actions through Dream API. In its technical explanation, Dream Labs describes a World Action Model that predicts video and actions together. For compatible bimanual robots with parallel grippers, the model can specify gripper poses, with each robot's inverse-kinematics solver translating those targets into joint commands.

That division leaves an important piece of execution with the machine. The same requested position can create different difficulties for arms with different reach, motion constraints or surrounding obstacles. Dream Labs says zero-shot performance varies across robots and depends partly on the solver, so compatibility needs to be evaluated on the equipment and tasks a builder actually intends to use.

The interface is commercially interesting precisely because it creates a place where different suppliers can cooperate. A hardware company can concentrate on the body while testing a model supplied by somebody else. The customer's eventual experience still spans both, bringing the quality of the connection into the product decision along with the quality of either component.

This also clarifies what the launch establishes. Dream Labs has made its model accessible and described a bounded route for transfer across hardware. The useful question for a developer is how that starting capability behaves on a specific job, including where demonstrations or additional engineering remain necessary. That question can now be investigated through the service rather than answered only by a selected launch clip.

From partner testimony to a developer's own evaluation

Dream Labs reports working with more than 20 hardware partners. In the funding announcement, MakerMods co-founder and CEO Ryan Chan describes a robot attempting previously unprogrammed tasks within minutes of using Dream-1. That is a named partner's experience, giving builders a concrete reason to investigate while leaving room for their own evaluation of repeatability and task performance.

Dream Labs also offers Fine-tuning as a Service, through which developers submit demonstrations for adaptation. The existence of that path is commercially revealing: a general model and a customized application can sit inside the same supplier relationship. A builder can begin with shared capabilities and then invest in the behaviors that make a particular installation useful.

For a small hardware maker, that arrangement changes what it needs to maintain internally. It may be able to draw on a model that improves beyond its own engineering team, while its customer relationship continues to include setup, execution and support. The appeal depends on how well those responsibilities are understood when the robot reaches a new task or environment.

Related DevCuration coverage of RSYNQ's physical-AI coordination software follows another part of this integration problem. Origin's construction robots show a more task-specific route into real work. The distinctions matter when assessing where a supplier expects to earn its place in a robotics installation.

Pricing the model beside the work

Dream Labs lists Dream-1 at $3/hour and Dream-1-flash at $2/hour for model operation. Those prices give builders a visible input to an experiment or service budget. Hardware, integration, supervision and the handling of unsuccessful attempts still affect the cost of the work a customer wants completed.

The technical documentation describes Dream-1 as a research beta and says it has not been validated for safety-critical use. That boundary belongs inside deployment planning, especially where robot motion brings people or valuable equipment into the setting. Developers need to design evaluations around the intended environment rather than borrow certainty from an unrelated demonstration.

The $90M financing supports a company making access to robot intelligence a product in its own right. Its public playground, API and adaptation service give outsiders several ways to explore that proposition, from a browser command to their own hardware. For Dream Labs, each new connection brings another developer's task into the process of deciding which capabilities are useful enough to keep calling.