Luma establishes Open Physical AI Lab, an open science initiative

Luma launches the Open Physical AI Lab, an open science initiative to solve the generalization crisis in physical AI by developing accessible World Models.

Luma is establishing the Open Physical AI Lab, a laboratory designed with an open science model. Its objective is to solve the generalization problem in physical AI—which is AI capable of acting in the real world—and to develop World Models to understand and interact with this environment.

The initial observation focuses on current robots, trained to perform precise tasks in precise environments from few examples. This lack of data, according to Luma, perpetuates a generalization crisis that the usual approach, collecting teleoperation data for every imaginable task, cannot solve due to a lack of physical and economic resources.

The company bases its approach on its own work: for four years, it has trained foundation models in 3D, image, video, and unified language-image generation, and built an infrastructure to leverage multimodal data at internet scale. From this, it derives the conviction that generalization is not exclusive to text and can extend to the physical world.

The lab plans to make the technical foundation it develops freely accessible, collaborate with academia and other laboratories, build progress and safety evaluations, and forge industrial partnerships concerning chips, hardware, and physical agents. This approach addresses, according to the company, the fear of a few players concentrating global intelligence and physical infrastructure.