Dream Labs has released Dream-1, a World Action Model that enables zero sample control of an unprecedented dual arm parallel gripper robot, performing operations in unseen tasks and environments, and offering services at a price of approximately $3 per hour through a public API; The company also announced a seed round financing of $90 million to exit stealth mode. ⁠ The main problem they solve is that the model can generalize zero sample information to unseen environments, tasks, and even other robots, so that the new type of robot can have practical capabilities from the beginning. How to achieve this depends on the technical documentation, https://dreamlabs.ai/research/dream-1/ How to use: The Dream-1 is designed to be as close to plug and play as possible. Just connect your robot, point it to the API, and it will run. If they claim to have achieved this without any background in machine learning, it is highly similar to the business model of the robot inference market that I shared earlier. Team background: Dream Labs was co founded by four core members from NVIDIA's GR00T robot research team in 2026. During their time at NVIDIA, they led the publication of the DreamZero paper introducing World Action Models, and subsequently left to form the company and launch Dream-1. ⁠ This was developed in Nvidia as Zero, which later became an upgraded version of One. CEO Joel Jang @ jang_yoel (Korean, former GR00T World Model Leader, leading LAPA, DreamGen, DreamDojo, and DreamZero); CTO Ge Yunhao (former GR00T video model leader, involved in NVIDIA Cosmos series). Third party evaluation of the model: Robocurve PBC, a third-party public welfare robot model evaluation agency, evaluates their model: https://(x.com)/robocurve/status/2112232494330483085 Previous research: The core paper of DreamZero is "World Action Models are Zero shot Policies", proposed by the Nvidia team. The World Action Model (WAM) is based on pre trained video diffusion backbone to jointly predict future videos and actions, achieving a generalization improvement of over 2 times for tasks/environments, 7Hz real-time closed-loop control, and cross ontology adaptation with only 30 minutes of data; DreamZero (Core WAM Paper): "World Action Models are Zero shot Policies" https://arxiv.org/abs/2602.15922 https://((((github.com))))/dreamzero0/dreamzero DreamDojo（《A Generalist Robot World Model from Large-Scale Human Videos》 https://arxiv.org/abs/2602.06949 https://((((github.com))))/NVIDIA/DreamDojo DreamGen（《Unlocking Generalization in Robot Learning through Video World Models》 https://arxiv.org/abs/2505.12705 https://research. (nvidia.com)/labs/gear/dreamgen https://((((github.com))))/nvidia/GR00T-dreams 《Latent Action Pretraining from Videos》 https://arxiv.org/abs/2410.11758 https://((((github.com))))/LatentActionPretraining/LAPA