Physical Intelligence's π0.7 Can Figure Out an Air Fryer From Two Training Examples
Physical Intelligence released π0.7 on April 16, 2026. The San Francisco startup's headline claim: the model can combine skills learned in separate contexts to solve tasks it was never explicitly trained on. They call this compositional generalization. The air fryer demo is their proof of concept.
The Air Fryer Experiment
π0.7's training dataset contained exactly 2 episodes involving air fryers. Despite that, the model produced a functional attempt at cooking a sweet potato with zero coaching. With step-by-step verbal instructions, it succeeded.
The early success rate was 5%. After roughly 30 minutes of prompt engineering, it hit 95%.
That gap deserves attention. A 5% baseline that requires 30 minutes of prompting to become useful is a different story than "robot figures out appliance from scratch." Both things are true simultaneously.
What It Can and Cannot Do
Physical Intelligence is upfront about one limitation: π0.7 cannot execute complex multi-step tasks from a single high-level command. That matters for anyone imagining a general-purpose kitchen robot responding to "make dinner."
On the benchmark side, π0.7 matched Physical Intelligence's own specialist models on coffee making, laundry folding, and box assembly. The comparison group is internal. Standardized robotics benchmarks do not exist yet, so external validation of these numbers is not currently possible.
The Company Behind It
Physical Intelligence is two years old. Co-founder Sergey Levine also holds a professorship at UC Berkeley focused on AI for robotics. Researcher Lucy Shi is a Stanford computer science Ph.D. student.
The company has raised over $1 billion and was last valued at $5.6 billion. Reports indicate a new funding round is in discussion that would push the valuation to approximately $11 billion.
Reading the Tea Leaves
The compositional generalization result is genuinely interesting as a research direction. A model that can combine narrow skills into novel task solutions would address one of the core limitations of current robotics systems. Whether π0.7 does this reliably at scale, or under what conditions it fails, is not answered by a sweet potato demo.
The funding trajectory suggests investors are not waiting for standardized benchmarks to make their decisions.
Source: Techcrunch