Nvidia GR00T researcher starts a company: raises $90 million, launches cross-platform robot AI.

Beating AI News Flash: Dream Labs, founded by four former researchers from Nvidia's GR00T project, has announced the completion of a $90 million seed round of financing and the opening of the API for its robot model Dream-1.

Developers only need to provide the robot's camera feed, current state, and task instructions, and the model can generate robotic arm movements. The standard version costs $3 per hour, and the lightweight version costs $2.

In the past, switching to a different robot often required re-collecting data and adapting the model. In February of this year, the team's DreamZero, launched by Nvidia, required 30 minutes of data from the target robot for additional training.

Dream-1 eliminates this step. It learns the relationship between video and actions through a world action model, predicts where the gripper should move, and then the robot's own control program executes it.

The team tested 27 robots not seen during training in a simulation environment, including the PR2 launched in 2008. Test partners from 8 countries also demonstrated real-machine operations such as stacking cups, folding paper boxes, and pulling tissues. However, it currently mainly supports dual-arm robots equipped with parallel grippers.

In real-machine evaluations by the third-party institution Robocurve, Dream-1 achieved an average task progress score of 59.3%, surpassing Google Gemini 4 Argon's 48.2%, with an estimated model cost per attempt of only $0.15, about 1/16 of the latter.

However, another 400 tests publicly released by Dream Labs show that the complete success rate for pick-and-place tasks is only 40%, and for tool use it is as low as 6%.

The model is still in the research and testing stage and has not yet passed validation for safety-critical scenarios.