Luma has announced a new research effort aimed at one of robotics' hardest problems: helping machines generalize beyond narrow, preprogrammed tasks. The company said it is launching the Open Physical AI Lab, an open science initiative focused on building world models for physical AI and robotics.
The lab is meant to address what Luma describes as a major gap between current robot systems and general-purpose AI models. While large language models and multimodal systems can handle a wide range of familiar and unfamiliar prompts, Luma says robots today are still trained to repeat specific actions in specific settings using relatively small amounts of data. That, according to the company, limits their usefulness in real-world environments.
Luma frames the issue as a data problem as much as a model problem. The company argues that the common answer, collecting ever more teleoperated demonstrations for every possible task and task combination, is not practical at scale. It says the economics and physics of that approach make it unsustainable.
The company points to its own work over the past four years on multimodal AI as the basis for the new lab. During that period, Luma says it trained foundation models across 3D, image, video and unified language-visual generation. It also built infrastructure for processing large-scale multimodal data and training generative models more efficiently. Based on that work, Luma says it has concluded that generalization is not limited to text and can also be learned from raw multimodal data in the physical world.
According to Luma, the Open Physical AI Lab will focus on research that can scale methods for generalization and help develop world models capable of understanding and interacting with the physical world. In the company's view, improvements in this area could lead to robot systems being used across factories, research labs, hospitals, streets and homes.
Luma also argues that physical AI could have broad economic effects, including easing labor shortages and helping address stagnation in global economies. The company says these systems may eventually become a core layer of production in many nations.
At the same time, Luma says it wants to avoid a future in which a small number of companies control the intelligence and physical infrastructure that robots depend on. The company warns that such concentration would pose risks to democratic freedom, and it compares that concern to the concentration of power it sees in parts of the large language model industry.
To counter that trend, Luma says the new lab will operate as an open science effort. The company plans to build foundational tools and make them available so other teams can use, adapt and build on them. It also says it will work with academic researchers and peer labs on core science, create benchmarks to measure progress and safety, and partner with industry on chips, hardware and physical agent systems.
The announcement positions the lab as both a research project and a broader infrastructure effort. Luma says its experience with multimodal systems and large-scale training makes it well suited to pursue the problem of physical AI generalization. The company is also recruiting researchers and inviting outside partners to work with the lab.
For now, the launch signals Luma's intention to move deeper into robotics research while keeping the effort open to other participants in the field. The company is betting that world models, open infrastructure and broader collaboration can help move physical AI beyond today's task-specific limitations.