South Park Commons uploaded a 42-minute conversation with Physical Intelligence co-founder Sergey Levine on Oct. 22, 2026. South Park Commons partner Aditya Agarwal asks him about the company’s robot foundation models, from π0 to π0.7.
Levine describes an earlier experiment at Google: Robots in a repurposed conference room practiced grasping objects. They got good at grasping but not at other tasks, he says. His conclusion was that robots need prior knowledge as well as experience. Agarwal frames Levine’s broader approach as working on many problems together rather than building a specialized system for each task. Levine puts the idea this way: For a difficult intelligence problem, “making the problem as big as possible actually makes it easier.”
That need for prior knowledge shapes Levine’s explanation of π0. A robot foundation model takes observations such as camera images and a task instruction, then produces continuous actions, which can include commands for a robot’s joints. Levine says a large language model is a useful starting point because it has knowledge but “is not grounded.” He calls π0 an “outline” to which reasoning and other data sources could be added.
For π0.5, Levine says the company added an internal reasoning component. Given an instruction to clean a kitchen, he explains, the model can break the job into steps such as picking up a dish or wiping a spill before translating those steps into actions.
Levine describes π0.7 as the most recent model Physical Intelligence had detailed and says it was designed for “zero-shot generalization out of the box.” He says it can tackle new tasks, offering the example of working out how to use an unfamiliar kitchen appliance. But he cautions that robotics is “not at the stage yet where you can just tell it exactly what you want, and it will do anything.”
Comments