Dream-1

Dream-1 is a World Action Model from Dream Labs, a San Francisco company, available through a public API. Per the launch post, it controls bimanual robots with parallel grippers that it has never seen, on tasks and in environments it was never shown, without per-robot fine-tuning. Dream Labs describes it as a research beta.

What It Is

Dream-1 is a robot foundation model, a "robot brain" that takes camera images, the robot's current state, and a text task, and returns actions for the robot to execute. Developers connect a robot to the API through a Python client, set observations, and request actions. Each call returns 1.6 seconds of actions. Two variants are described: Dream-1 and the faster, lighter Dream-1-flash.

How Zero-Shot Control Works

Instead of outputting joint angles tied to one body, Dream-1 can predict end-effector poses: where each gripper should be and whether it is open or closed. The robot's own inverse kinematics solver converts those poses into joint commands. The model sees through a head camera and a wrist camera on each arm. As a World Action Model, it predicts video and actions together, so different robot bodies appear to it as pixels in a scene. Dream Labs reports zero-shot results on 27 simulated unseen robots, measured relative to YAM, the robot it was trained on, plus real-robot demonstrations.

Fine-Tuning and Limits

For robots or tasks where zero-shot is not enough, Fine-tuning-as-a-Service adapts the model from the user's own demonstration data. The post shows examples with 5 minutes of data on some robots and more for embodiments far from the training data, such as dexterous hands. Documented limitations include dependence on IK solver quality, no control over how the arm reaches a pose, and possible infeasible predictions for robots with very different kinematics. Dream Labs states Dream-1 is a research beta that has not been validated for safety-critical use and should be tested in controlled settings.

Current Scope

Today Dream-1 is a manipulation policy for bimanual robots with parallel grippers. Dream Labs says it hopes to extend it to mobile manipulators and dexterous hands. A Robot Playground lets users prompt a real robot without owning one.