A home robot that can handle many ordinary chores has been a long-running science fiction promise. Physical Intelligence, a San Francisco startup also known as PI or π, has shown a step toward that idea with a single artificial intelligence model trained to control robots across several domestic tasks.
The model, called π0 or pi-zero, is not being presented as a finished household helper. Its importance is narrower and more concrete: it suggests that robotics may be moving toward the kind of general-purpose learning that made large language models useful across many text-based tasks.
What Physical Intelligence demonstrated
Physical Intelligence has spent the past eight months developing π0, a foundation model for robots. The company trained it on huge amounts of data from several kinds of robots performing household chores, often using humans to teleoperate the machines and provide examples for the system to learn from.
Videos released by Physical Intelligence show different robot models performing practical work around the home. A wheeled robot reaches into a dryer and retrieves clothing. A robot arm clears a table covered with cups and plates. A pair of robot arms grabs and folds laundry. Another robot builds a cardboard box by bending its sides and fitting pieces together.
Those tasks matter because they demand more than simple repetition. Folding clothing, for example, requires handling flexible objects that change shape, crumple and move unpredictably. Karol Hausman, the company’s CEO, says this makes clothing a difficult test of general physical understanding.
The model also shows behaviors that look familiar from human problem-solving. In the examples described, it shakes T-shirts and shorts so they lie flat before folding them.
Why this approach is different
Robots in factories and warehouses are often powerful but narrow. Many follow precisely choreographed routines and have limited ability to understand their surroundings or adjust when conditions change. Even robots that can see and grasp objects usually remain limited in dexterity and range of use.
Physical Intelligence is trying to change that by applying a lesson from modern AI. Large language models became more capable by learning from enormous bodies of text. The company’s goal is to build a comparable kind of general capability for the physical world, using robotic data instead of internet text.
Hausman describes the company’s method as a general recipe that can use data from many embodiments and robot types, similar in spirit to the way language models are trained. Sergey Levine, a cofounder of Physical Intelligence and an associate professor at UC Berkeley, says the amount of data used is larger than any robotics model ever made, by a very significant margin, to the company’s knowledge.
Levine also sets expectations carefully. “It’s no ChatGPT by any means, but maybe it’s close to GPT-1,” he says, referring to the first large language model developed by OpenAI in 2018.
The data problem for home robots
The comparison with language models also reveals the central obstacle. Text data exists at massive scale. Robot training data does not. A robot learning system needs examples of machines acting in the physical world, and those examples are harder to collect.
That is why Physical Intelligence has to generate its own data and develop methods that can learn from a more limited dataset than the one available to language models. To build π0, the company combined vision language models, which are trained on images and text, with diffusion modeling, a technique borrowed from AI image generation.
The broader idea is that learning should transfer across tasks and robot bodies. Earlier robot training efforts often focused on one machine learning one task, because researchers treated the knowledge as difficult to reuse. More recent academic work has shown that transfer can become possible with enough scale and fine-tuning.
One example in the source is Open X-Embodiment, a 2023 Google project that shared robot learning between 22 different robots at 21 different research labs.
Useful progress, not a finished helper
π0 is still imperfect. Hausman notes that the system can fail in surprising and amusing ways, much as modern chatbots sometimes do. In one case, a robot asked to load eggs into a carton overfilled the box and forced it shut. In another, a robot suddenly threw a box off a table instead of filling it.
Those failures are important because they show the gap between demonstration and dependable deployment. A home is full of variability: cluttered surfaces, soft objects, fragile items and tasks that change from one moment to the next. A robot that works in that setting needs more than a memorized routine.
The commercial incentive is clear. More capable robots could take on a wider range of industrial tasks, possibly after only minimal demonstrations. They would also be better suited to homes, where messiness and variation are unavoidable.
Interest in artificial intelligence has already increased optimism about robotics. Tesla is developing a humanoid robot called Optimus, and Elon Musk recently suggested it would be widely available for $20,000 to $25,000 and capable of doing most tasks by 2040.
What π0 signals about the future of robotics
The strongest claim supported by Physical Intelligence’s work is not that general home robots have arrived. It is that robotics researchers may now have a more scalable path toward them.
π0 shows that one AI model can guide multiple robot types through multiple useful chores. It also shows that the path depends on data, transfer learning and techniques that connect vision, language and physical action.
Levine frames the work as an early structure rather than a final system. “There's still a long way to go, but we have something that you can think of as scaffolding that illustrates things to come,” he says.
For now, the value of π0 is in what it makes visible. A robot that unloads a dryer, clears a table, folds laundry or builds a box is no longer only an image from old television science fiction. It is also a research direction being tested with foundation models, large-scale robot data and the hard physics of ordinary homes.