Physical Intelligence, a 2024 San Francisco startup, is the cleanest current test case for the bet that one big AI model can learn many physical tasks the way large language models learned to handle many text tasks.
The robot's hand hovered over the espresso portafilter, gripped it, locked it into the group head, and started the shot. A few feet away, another arm folded a t-shirt into thirds, smoothed the collar, set it down, and reached for the next one. Down the bench, a third peeled a potato in a single, unbroken spiral. None of these were separate programs stitched together for a press demo. According to New Scientist's profile of Physical Intelligence, they were the same control system, running on the same model, on machines that were never designed to be the same kind of machine.
That distinction, one brain, many bodies of work, is the spine of the bet Physical Intelligence is making. The company, a San Francisco robotics startup founded in 2024, is building what its founders call a general-purpose robotic brain: a single policy that can be ported across different hardware and taught new skills without being rebuilt from scratch for each one. The demos on show in the company's warehouse, which included coffee, laundry, vegetable prep, and kitchen cleanup, were presented as evidence that the same stack could move between them.
The pattern is borrowed, almost wholesale, from the last AI wave. Large language models (LLMs) such as the systems behind ChatGPT replaced a zoo of task-specific tools with one model trained on a broad corpus. Physical Intelligence, and a small group of academic and industrial labs working on the same idea, are trying to do the same thing for physical action, using what researchers call vision-language-action (VLA) models: neural networks that take in camera images and instructions, and output motor commands. The pitch is the same one that worked for text. If you train one big model on enough variety, it generalizes to tasks it was never explicitly taught.
This is not how industrial robots have historically been built. A factory arm that welds a car door is a frozen program: every motion is taught, every exception is a fault. A Tesla humanoid or a Boston Dynamics humanoid, by contrast, is built around the assumption that one body will do many things. Physical Intelligence's wager is a third path: stop tying the intelligence to the body at all. Teach a model to fold a shirt, and let the same model, with the same weights, drive a different arm in a different kitchen tomorrow.
Whether that actually works outside a curated warehouse is the part the company cannot yet answer, and the part the field's own skeptics are quick to name. Both Berkeley roboticist Ken Goldberg and Oxford's Michael Wooldridge, quoted in the New Scientist piece, frame general-purpose robotics as a goal that has been on the field's whiteboard for decades, and VLA models as a hopeful but unproven translation of LLM enthusiasm into a domain where the cost of being wrong is a dropped mug, not a hallucinated paragraph. The model that folded the shirt did so on hardware the company controls, in a room the company lit, in a sequence the company chose to show.
What the startup is really selling, then, is a research bet with a clear thesis and a clock. If one model can absorb enough variety, the same generalization trick that took chatbots from toys to tools could take robots from single-purpose machines to something closer to a household appliance you can talk to. If it cannot, the gap between a curated demo and a kitchen at 7 a.m. on a Tuesday will be the same gap it has been for forty years, and the next AI wave will have washed up at the edge of the screen.