Dream Labs, a startup founded by former NVIDIA robotics researchers, has raised $90 million in seed funding and launched a public API for Dream-1, a model designed to carry skills across different robots, Unite.AI reported on October 19, 2026. The round was led by B Capital alongside Kindred Ventures, with participation from A.Capital, Unusual Ventures and other investors. The company says the money will mainly pay for compute as it expands model capabilities and capacity.

The problem Dream Labs is going after

Robotics has a costly bottleneck: a machine that learns a task does not automatically pass that knowledge to a different machine. A policy trained on one arm usually does not work on another with different joints, grippers or camera placement, so teams often collect new demonstrations for each body. Dream Labs wants developers to start from a shared model and adapt it only when a particular robot or application needs more specialized training.

The launch includes an API, a fine-tuning service and a browser-based Robot Playground that lets users send instructions to physical robots in the company’s San Francisco office. Models run through Dream Labs’ hosted service, so developers do not have to build and operate a robotics foundation model themselves.

How Dream-1 works

Dream Labs calls Dream-1 a World Action Model. According to the company’s technical account, as summarized by Unite.AI, it predicts video and actions together: anticipating how a scene will change is linked to deciding what the robot should do next. For background on related approaches, see our explainer on vision-language-action models in robotics and the glossary entry on world models.

The key design choice is the output. Because joint commands that work on one arm will not necessarily work on another, Dream-1 predicts end-effector poses instead: the desired position and orientation of each gripper, and whether it should be open or closed. Each robot’s own inverse-kinematics solver then converts those targets into movements its joints can execute. That gives different machines a common interface without requiring identical hardware.

The model uses three camera views, one of the scene and one on each wrist. In the developer loop, the robot supplies images, its current state and a task instruction, and the model returns actions.

What the claims do and do not show

The company’s zero-shot capability currently focuses on bimanual robots with parallel grippers, meaning two-armed machines with simple open-and-close grippers. Zero-shot means the model attempts tasks on an unfamiliar robot without first training on demonstrations from it. Dream Labs’ published evaluation covered 27 unseen robot embodiments in simulation, with 300 rollouts per embodiment, and results were reported relative to YAM, the robot on which Dream-1 was trained.

Two cautions follow. Performance varied substantially across hardware, so the headline claim of cross-robot transfer is more useful read alongside the detailed results. And relative scores should not be confused with absolute task-success rates. The research page also distinguishes full completion from partial progress. The control interface has trade-offs as well: predicting where a gripper should go does not fully determine the path the arm takes, and obstacles, joint limits and solver quality can affect performance. The evaluation was in simulation, not on a broad set of physical robots.

Why it matters

If transfer works reliably, the economics of robotics change: data collected on one platform becomes useful for many, and smaller companies could build on a shared model rather than training from scratch. That is the same logic that made general language models valuable. The open question is whether physical variety, from friction to sensor noise, can be bridged as neatly as the demonstrations suggest. For context on robots entering industry at scale, see our report on Amazon’s one million robots.

What to watch next

  • Independent evaluations of Dream-1 on physical robots rather than simulation.
  • Support beyond bimanual parallel-gripper machines, such as dexterous hands or mobile bases.
  • Pricing and usage terms for the API and fine-tuning service.
  • Whether competing robotics foundation-model companies adopt similar end-effector interfaces.

Key takeaways

  • Dream Labs raised a $90 million seed round led by B Capital and Kindred Ventures.
  • Dream-1 predicts video and gripper poses together and relies on each robot’s inverse kinematics.
  • Zero-shot transfer is limited to bimanual parallel-gripper robots so far.
  • Reported tests covered 27 unseen embodiments in simulation, with varied results.
  • A public API, fine-tuning service and Robot Playground are available now.

Frequently asked questions

What is a World Action Model?

It is Dream Labs’ term for a system that predicts how a scene will change and what actions a robot should take, in one model.

What does inverse kinematics do here?

It is the calculation that turns a target gripper position into the joint movements a specific robot arm needs.

Does this mean any robot can use Dream-1?

No. The current zero-shot scope is two-armed robots with parallel grippers, and performance varied across tested hardware.