Dream-1: An AI capable of controlling robots it has never encountered

Dream Labs introduces Dream-1, an AI capable of controlling new robots without specific training, accessible via API starting at $3 per hour.

A dual-arm robot equipped with grippers could soon perform its first tasks without having been trained on its own hardware. This is the proposal of Dream-1, an artificial intelligence model developed by Dream Labs capable of transferring its knowledge across different machines. In its demonstrations, the system grabs objects, moves them, stacks containers, or manipulates accessories based on natural language instructions, including on robots absent from its training data. The American startup is accompanying its exit from stealth with a $90 million seed round and the launch of a public API, priced starting at three dollars per operating hour. A Single Model for Multiple Robots

Robotics suffers from a recurring difficulty: skills acquired by one machine are difficult to transfer to another. Two robots may each possess two arms and grippers, but their geometry, joints, and control systems differ. Developing a new capability often requires collecting demonstrations on the hardware in question, then training or fine-tuning a model. This is an expensive operation, particularly for small-series manufacturers and teams lacking significant research resources.

Dream Labs seeks to reduce this dependence on machine-specific data. Its model belongs to the World Action Models (WAM) family, introduced by the DreamZero work in February 2026. These systems learn to associate visual observations, the probable evolution of a scene, and the actions that produce that evolution. The ambition is to have a model that can recognize a task to be performed and control a robot different from those used during its training.

This capability, called zero-shot generalization, constitutes the main argument for Dream-1. A developer with a compatible robot can send it camera observations and an instruction, then receive motion commands without first conducting training on demonstrations from that robot. The company nevertheless restricts this promise to a specific hardware category: bimanual robots equipped with parallel grippers. It is not yet a model capable of controlling just any machine indiscriminately. Twenty-Seven Unknown Robots Subjected to the Same Evaluations

To measure this transfer capability, Dream Labs tested Dream-1 on 27 novel robotic configurations in simulation, ranging from relatively old research arms to newer equipment. The evaluated machines include PR2, Baxter, ALOHA, Unitree R1, OpenArm 2.0, and configurations using two UR5e arms. The model was also demonstrated on several physical robots in laboratories and at international partners, including Hugging Face in Paris, KAIST in Seoul, AgiBot in Shanghai, and Scale AI in San Francisco.

Results vary significantly depending on the hardware. In simulation, Dream Labs reports a relative score of 91.2% for PiPER, compared to 82.8% for Trossen Workbench, 73.8% for PR2, 63.9% for Unitree R1, and 54.2% for SO-101. The YAM robot, used as a reference because it was part of the training, corresponds to 100%. These values express performance relative to this benchmark, rather than absolute task success rates.

Each simulated configuration undergoes 300 trials, including tests of instruction following and generalization to unseen tasks. Physical demonstrations include putting objects away in containers, stacking cups, manipulating bottles, and folding a box. Some videos show only a single trial, while others show one success among several attempts: they thus illustrate possible behaviors without establishing uniform reliability across the different robots. Predicting the Scene and the Actions Required to Achieve It

The design of Dream-1 relies on a shared representation of actions. Traditional robotic models can directly predict joint angles, making them tightly bound to a machine's mechanics. Dream-1 can use this method for the robots it natively supports, but it can also output targets describing the position and orientation to be reached by each gripper, as well as its open or closed state. An inverse kinematics solver specific to the robot then translates these targets into joint movements.

This separation allows multiple machines to share a control interface despite their physical differences. Dream-1 also analyzes observations coming from three cameras: a overhead view and a camera on each of the two wrists. The model learns to jointly predict future images and actions in order to link the expected visual outcome to the movements required to achieve it.

The method nevertheless has limitations. Accuracy depends on the quality of the inverse kinematics solver, while the model does not directly control all the details of the trajectory followed by the arms. Certain movements can thus become difficult near obstacles or joint limits. A machine whose geometry or physical capabilities diverge sharply from those encountered during training may also receive targets that are impossible to execute. Performance Evaluated Against Generalist Models

Dream Labs also relies on an evaluation conducted by Robocurve, an organization specializing in benchmarking robotic capabilities. On RoboDojo-RC Tier 1, which evaluates physical manipulation tasks, Dream-1 achieves an average progress score of 59, compared to approximately 48 for Gemini 4 Argon, the best competing result indicated in the comparison. This is a progress score over attempts, not the percentage of fully completed tasks.

According to the report cited by Dream Labs, the estimated cost per attempt is sixteen times lower than that of Gemini 4 Argon, while trial duration is 1.7 times shorter than that of GPT-6 Astra, presented as the fastest generalist language model in this evaluation. These comparisons apply to the specific configurations of RoboDojo-RC Tier 1 and do not constitute a universal ranking of all artificial intelligence models.

Dream-1's positioning is also different from that of a conversational assistant. Where a large language model can devise a plan and select a sequence of steps, Dream-1 seeks to provide