Luma AI Targets Robotics Generalization With Open Physical AI Lab

Luma AI is launching an open physical AI lab to pursue a specific goal: robots that can generalize, rather than repeat tasks they have been individually shown. Amit Jain frames the problem as a gap between how large language models are used and how robots are typically trained. In his account, “pretty much all robots” are still trained by being shown a few examples of specific tasks. The field, as he puts it, trains “one task at a time.”

That is the constraint Luma is trying to attack. Jain contrasts today’s robotics practice with large language models, where a large model can be trained and then applied to a wide range of tasks, including problems it has not seen before. For robotics, he says, that ability remains missing. The practical ambition is not a robot that can repeat one demonstrated behavior, but one that can handle a new instruction in a new situation: do this, then when that is done, go take care of something else, or eventually “run my house.”

Ed Ludlow describes the problem as generalization: a robot may be presented with a new scenario for the first time. Jain agrees. Generalization, in his definition, is the problem of allowing robots to solve generally any task, rather than only the narrow tasks for which they have been explicitly trained.

Bloomberg’s on-screen fast-facts graphic identifies Luma AI as founded in 2021, headquartered in Palo Alto, California, and developing multimodal AI models for video. That background matters to Jain’s argument because he does not present physical AI as a separate robotics niche. He presents it as a continuation of work on large multimodal systems, extended into control, simulation, and embodied action.

Even with access to large amounts of synthetic or virtual data, the technical tension remains physical grounding: models need a connection to “the real physics of the world.” Jain’s answer separates into two parts: what Luma is working on technically, and why the work should be open.

On the technical side, he argues that the current best approach is essentially brute force: collect data for every task and every combination of tasks one can imagine. His examples range from picking up cups to digging mines. If robotics remains dependent on task-specific data collection, the path to general-purpose physical intelligence requires a practically impossible catalog of human activity.

Jain says Luma’s work over the past four years has been in building “general systems out of multimodal data”: internet-scale multimodal data from which the company tries to extract signals that allow control, simulation of reality, and physical control. The lab’s job, as he describes it, is to bring that capability into physical AI.

The claim is not that Luma has already solved general-purpose robotics. Jain’s case is that the company’s experience with video, image, and 3D models is relevant because physical AI will require large-scale multimodal data infrastructure, not only one-task-at-a-time demonstrations.

Jain insists that the “open” part is not incidental. When Ed Ludlow asks whether open means open source, Jain says it means both open science and open source. He then explains openness through the scope of where he expects physical AI to be deployed.

Physical AI, in his view, will not be confined to software interfaces. It will be in houses, factories, food systems, hospitals, scientific labs, and streets. Robots will manufacture “everything we depend upon” and participate in the systems that produce what people eat and use. Because of that scope, Jain says it is “completely untenable” for one or two actors to control the entire stack.

That is the core argument behind the lab’s open initiative. Jain says Luma wants small groups to be able to take these technologies and build them into productive systems. Openness, in this framing, is a way to avoid a world in which the physical AI stack is controlled by only one or two companies.

Caroline Hyde challenges him on whether this is mainly a philosophical position. She points to Meta’s shifting posture on open source as part of that challenge, and raises the commercial and geopolitical pressures around funding, China, and national tensions.

Jain accepts the philosophical label but argues that philosophy is required because physical AI is not “just a tool.” AI is already more impactful than many expected two years earlier, he says, and physical AI will be deployed even faster because of its economic impact.

His answer to the funding question is that concentrated control over the means of production is not economically stable anyway. Nations, he argues, will not be comfortable with one or two companies outside their borders controlling their means of production. He therefore presents openness not just as a moral stance but as an economically sound one.

The economic model he sketches is an ecosystem: chip partners, “model brain” providers such as Luma, and deployment partners working together to turn physical AI into systems of productive work. Jain says current intelligence systems, especially LLMs, are moving in the “wrong path” on this question of concentration.

During this portion, Bloomberg shows Luma-sourced AI-generated video imagery and captions it: “JAIN: PHYSICAL AI NEEDS TO TRAIN ON OPEN-SOURCE DATA.” The caption reinforces the same point Jain is making in the exchange: for him, the openness of the data and research stack is part of the physical AI project, not a separate branding choice.

Caroline Hyde asks why Luma is the right company to pursue this when there are other efforts in the field, naming World Labs and Fei-Fei Li. Jain’s answer is that physical AI will not be solved primarily by traditional domain expertise alone. He draws an analogy to language models: language models, he says, were not solved by linguists.

For Amit Jain, the relevant capability is large-scale multimodal data infrastructure. He describes that as Luma’s “bread and butter.” The company, he says, has produced strong models in 3D, images, and video using raw internet data. That skill — turning large, messy multimodal data into model capability — is what he argues is essential to solving physical AI.

Jain’s funding argument returns to the same premise as his openness argument: physical AI will touch productive systems that countries and industries depend on, so the stack cannot plausibly be concentrated in one or two companies. He presents an ecosystem of chip partners, model providers, and deployment partners as the economic structure that should carry the technology into use.