Dream Labs exits stealth with Dream-1, a model that runs new bimanual robots with no fine-tuning

Dream Labs has emerged from stealth with $90 million in seed funding and the launch of Dream-1, a robot-control model the company says can operate unfamiliar two-armed robots on tasks it has never encountered, without any fine-tuning. The model is available worldwide through a public API, and the announcement quickly reached the front page of Hacker News.

A policy with no per-robot training

According to Dream Labs, giving Dream-1 a new bimanual robot with parallel grippers and connecting it to the API is enough for the robot to begin performing tasks immediately: no fine-tuning, no demonstrations collected from that specific machine, and no machine learning background required. The integration on the robot side is intentionally minimal — send observations to the API, receive actions back.

To support the claim, the company published rollout videos from beta testers across the United States, Europe and Asia, including Hugging Face, Scale AI, KAIST, AgiBot, BitRobot, Cortex and RoboCurve. The tasks shown include sorting candy into containers, pulling a single tissue from a box, stacking one cup on another, folding a cardboard box, standing three bottles upright and building a tower of cubes. Dream Labs describes these rollouts as autonomous, performed on tasks the model was not trained on.

The company also says Dream-1 transferred to robot hardware it had never seen during training, including the PR2, a bimanual research platform released 18 years ago, with no embodiment-specific data involved. Dream Labs considers this among the first reported cases of a model generalizing zero-shot to entirely unseen robot bodies.

Targeting robotics' cold-start problem

The launch post is built around an old dilemma in robotics: a robot needs to be useful before anyone will buy it, but making it useful has required data from that exact robot, which only exists once enough people use it. Self-driving cars escaped this loop because vehicles were useful with zero autonomy and generated training data through everyday driving.

Dream Labs argues two changes now break the loop for robots. First, hardware has become cheap: a capable two-arm setup costs about $2,000, the company points to seven semi-humanoid robots available for under $15,000, and Hugging Face's $399 Microduck — a robot that walks, waddles and roller-skates — sold more than 10,000 units within four days of its launch in late August, according to Dream Labs. Second, a model that generalizes zero-shot to unseen environments, tasks and robots lets new hardware be useful immediately. Together, the company argues, these enable a flywheel: cheap robots spread, useful robots generate data, and that data improves models for the next wave of robots.

Built on World Action Models

The research behind Dream-1 dates to February 2026, when the team published DreamZero, the paper that introduced World Action Models, a paradigm Dream Labs says the community has since embraced. The pivotal result came from post-training DreamZero on just 30 minutes of data from a YAM robot, after which it generalized zero-shot to tasks it had never seen — a degree of transfer that previously demanded hundreds or even thousands of hours of data specific to one robot body. That finding motivated the company's founding, and after roughly six months of building the team produced Dream-1.

Company-reported, with known limits

The results so far come from Dream Labs itself rather than independent evaluation, and the company acknowledges variability: zero-shot performance depends partly on the quality of a robot's inverse kinematics solver, for which it published a tuning guide, and outcomes differ from robot to robot. Dream Labs is inviting anyone with a bimanual, parallel-gripper robot to try the API and report back.

Why it matters

Per-embodiment data collection and in-house ML expertise are among the largest hidden costs in deploying learning-based robots today. If Dream-1's claims hold up, that cost collapses into an API call, opening serious robot deployment to small teams, researchers and prosumers who could never fund a data-collection effort of their own. Combined with sub-$2,000 bimanual hardware, a general policy could start the data flywheel the robotics industry has chased for years. The public API also means the claims are testable now by anyone with compatible hardware, and the beta tester list already spans three continents.