Inside Physical Intelligence, where robots learn like ChatGPT
-

From the street, Physical Intelligence’s San Francisco headquarters is almost invisible, marked only by a subtly colored pi symbol on the door. Inside, however, the space is alive with activity. Robotic arms sit atop long wooden tables, attempting everyday tasks like folding clothes, turning shirts inside out, and peeling vegetables.
The startup, co-founded by UC Berkeley professor Sergey Levine, is building general-purpose robotic intelligence — an approach Levine describes as “ChatGPT, but for robots.” Data collected from robots operating in labs, kitchens, warehouses, and homes is used to train foundation models that can generalize across tasks and environments.
The hardware itself is intentionally inexpensive and unpolished. The company’s bet is that strong intelligence can compensate for weak hardware, allowing robots to master mundane tasks that once seemed out of reach.