Dream-1

Dream Labs presents Dream-1, a World Action Model intended to let robots perform useful tasks without first collecting robot-specific training data. Its central argument is that robotics can escape a cold-start loop: cheaper hardware puts more robots into homes and small teams, while a model that generalizes across robots and tasks can make them useful from the start. The company points to two enabling developments: bimanual robot setups now costing around $2,000 and low-cost consumer robots selling at scale. More capable robots could generate real-world data that improves future models. Dream Labs says an earlier experiment - adapting its DreamZero model with 30 minutes of data from one robot - led to performance on tasks it had not been shown, motivating this broader effort.

Dream-1 is offered through a public API, reportedly for about $3 per hour, and is designed to work without fine-tuning or machine-learning expertise. Dream Labs reports tests on 27 unseen robot types in simulation and demonstrations across numerous real bimanual, parallel-gripper robots, including the 18-year-old PR2. It also describes task demonstrations from beta testers at 12 sites in multiple countries, alongside internal tests. Examples include sorting objects, folding fabric, moving items between containers, and operating drawers. These results are presented as early evidence of zero-shot generalization to unfamiliar embodiments, environments, and tasks - not proof of universal capability. Performance depends in part on each robot’s inverse-kinematics solver, and the demonstrations focus on bimanual robots with parallel grippers.