Dream-1

Dream-1, a new bimanual robot with parallel grippers, has been developed by Dream Labs, enabling it to perform tasks out of the box without any fine-tuning or machine learning expertise required. This development marks an important step towards general-purpose robot intelligence. Dream Labs has recently emerged from stealth, securing $90M in seed funding and making Dream-1 available globally through a public API.

In the robotics industry, a chicken-and-egg problem has persisted for decades: robots need to demonstrate usefulness to attract consumers, but their usefulness relies on data from their own operations, which is generated only when enough people start using them. To break this loop, Dream-1 represents a significant step.

Similar to how self-driving cars provided immediate utility with zero autonomy, allowing millions of people to drive and generate data for increasingly capable models, robotics now faces a different challenge. While robot hardware is becoming increasingly affordable, allowing more people to own robots, these robots often require extensive data and machine learning expertise to function effectively.

Dream-1 aims to bridge this gap by leveraging World Action Models (WAMs), a promising paradigm in robot learning. Initially introduced in a 2026 research paper, WAMs have shown the ability to generalize zero-shot to tasks unseen during training, often achieving this in just 30 minutes of data. Connected to Dream-1, this breakthrough enables it to control unseen robots for unseen tasks in unseen environments, all without requiring extensive embodiment-specific data.

By making Dream-1 accessible through a public API, Dream Labs aims to democratize generalizable robotics, enabling anyone with a robot to utilize its capabilities without needing advanced machine learning knowledge. Zero-shot generalization works by leveraging the shared EEF action space of bimanual robots with parallel grippers, allowing Dream-1 to issue joint commands to any supported robot despite its unique design.

This approach has demonstrated successful zero-shot performance on various bimanual, parallel-gripper robots, including the original PR2, a 18-year-old general-purpose bimanual research robot.