TECHNOLOGY · APR 16, 2026
Physical Intelligence robot models clean unseen homes and try chores they never practiced
pi0.5 tidied kitchens and bedrooms in homes outside its training data in April 2025. pi0.7, a year later, mixes skills it already has to attempt new tasks.
By Riley Hart · 3 sources linked below
Nothing here is for sale to households, and Physical Intelligence does not make a home robot. The research matters because a robot that works in a house it has never seen is what a general home helper needs. Expect these ideas to show up inside other companies' robots first, and judge any such robot on independent tests, not lab videos.
Physical Intelligence, a San Francisco robotics AI startup, builds the software brains that control robot arms. Two of its releases mark progress on a basic problem for home robots: working in a house that looks nothing like the training data.
On April 22, 2025 the company introduced pi0.5. It ran a wheeled robot with two arms that cleaned kitchens and bedrooms. The company says every test home was left out of the training data. Tasks included making the bed, putting clothes in a laundry basket, putting dishes in the sink, storing items in drawers and wiping a spill with a sponge. Training mixed about 400 hours of mobile robot data with web data and data from other kinds of robots.
The company's numbers are its own and have not been checked by outside testers. It reports an average task success rate of 94% in the unseen test settings and 83% in settings similar to training. It also says results get close to a model trained on the test home once training covers about 100 different homes. Physical Intelligence wrote that the model is "far from perfect" and often makes mistakes, both in planning and in movement.
On April 16, 2026 the company released pi0.7. The headline claim is that it can combine skills it already has to try tasks it was not trained on. In one test, a mobile robot was asked to load a sweet potato into an air fryer. On its own, the robot made a reasonable attempt but did not finish. When a person coached it step by step in plain language, it did much better.
In another test, a two arm UR5e robot folded laundry even though no laundry data had been collected on that robot. The Decoder reports an 80% success rate on t-shirts, which the company says matches skilled human operators using that robot for the first time.
The Decoder also raised a fair caution. The training data held two clips of a robot closing an air fryer and similar drawer opening clips, so the air fryer task may be less new than it looks. The company itself calls its skill combining results early.
Neither model is a product you can buy. These are research models shown on lab and test robots, and Physical Intelligence does not publish how often they fail in a full day of real chores. The progress is real on the company's own tests, but a dependable helper that tidies any home without supervision is not here yet.