Luma launches open physical AI lab
Luma, which builds video and multimodal generative models, is starting an open science lab to tackle generalization in robotics, arguing that the same internet-scale multimodal data can train physical AI.
Luma has announced the Open Physical AI Lab, a new research effort aimed at what the company calls the "generalization crisis" in robotics. It will research and scale methods for generalization and build world models for understanding and interacting with the physical world.
Luma's case is that physical AI is still in a "pre-AI-scale era." Robots are trained on small sets of examples and can only replay specific tasks in specific environments. The company calls the usual fix, scaling up teleoperation to collect data for every task, "physically and economically impossible."
Instead, Luma is betting on what it already knows. Over four years it has trained foundation models for 3D, image, video and unified language-visual generation, and built infrastructure to process internet-scale multimodal data. It says recent results from its Unified Models work have given it a "strong scientific conviction" that "generalization is not unique to text," meaning raw multimodal data can teach models physics, perception and interaction.
Luma presents openness as a matter of principle. It warns that letting one or a few companies control the world's "intelligence and physical infrastructure" is "a profound threat to humanity and democratic freedom." It says LLMs went that way and physical AI is heading the same way now. The company says it will release the substrate it builds so others can use and modify it. It also plans to work with peer labs and academia on core research, build evaluations for progress and safety, and partner with industry on chips, hardware and physical agent systems.
The announcement does not name any models, datasets, partners, timelines, licensing terms or funding for the lab.