Dream Labs Raises $90M, Offers Robot AI by the Hour

SAN FRANCISCO -- Dream Labs has raised $90 million in seed funding and launched an access to its robot AI models starting at $2 an hour, offering an alternative to building and operating their own AI systems.

The startup, founded by former Nvidia robotics researchers, announced the financing Oct. 19 of Dream-1, a robot foundation model designed to control unfamiliar machines .

B Capital led the round alongside Kindred Ventures. Other investors weree A.Capital, Unusual Ventures, SV Angel, Wing VC and Scribble VC.

Dream Labs targets a persistent problem in robotics, which is AI training on one machine often needs additional data and engineering before operating another one. The issue persists even when the robots perform similar tasks.

The approach called a World Action Model combines predictions of how a scene will change with the actions needed to produce those changes. The system uses visual information to help determine how different machines should manipulate objects instead of learning only the movements of a particular robot,

Developers connect their robots to Dream-1 through the company's cloud service. It supplies camera observations and instructions and the model returns actions that the robot executes.

ONE MODEL, DIFFERENT MACHINES

Dream Labs says Dream-1 can perform some manipulation tasks on previously unseen robots without additional training. The current system reports on two-armed robots with grippers. The company demonstrated different machines handling objects, moving containers and performing other tabletop tasks.

It also reported simulation tests involving 27 previously unseen robot configurations, and performance varied considerably. Dream-1 robot achieved 91.2% performance on a two-arm PiPER configuration, 73.8% on a PR2 and 53.9% on OpenArm 2.0. The reported figures are relative benchmark scores, not task-completion rates.

Dream Labs also said that for unfamiliar manipulation tasks, the reported success rate was 38% for pick-and-place operations, 18% for cloth and soft-object handling and 11% for insertion and fitting tasks.

The company claims performance also depends on each robot's inverse kinematics software, which translates desired gripper positions into joint movements. Robots with different movement parameters may find it hard to execute the modelled actions.

ROBOT INTELLIGENCE BY THE HOUR

Rather than requiring developers to train and operate their own foundation models, Dream Labs provides access to Dream-1 through a cloud-based application programming interface, or API, charging according to usage.

The flagship Dream-1 model costs $3 per hour, while the lighter Dream-1-flash version costs $2 per hour. Each API call returns approximately 1.6 seconds of robot actions. The hourly charges cover access to the AI model, not rental of a physical robot or the full cost of operating one. Hardware, electricity, maintenance, network connectivity and supervision will be charged separately.

Dream Labs says its approach could cheapen robot development for smaller companies and research teams that do not have sufficient resources to collect extensive training data. The startup traces its technology to DreamZero, a research project released in February 2026. That work explored whether models predicting both video and actions could transfer learned manipulation behavior between robots.

Dream-1 extends the approach into a publicly accessible service, although its current capabilities remain limited to certain types of manipulation. The company describes the release as a research beta and warns that the model can make mistakes. It has not been validated for safety-critical applications and should be tested in controlled environments, particularly when robots operate near people.

Dream Labs has not disclosed the number of paying API customers or demonstrated sustained commercial operations using Dream-1.