Traditionally, robots have been masters of the specific, bound by the rigid scripts of their training data. If a machine was taught to sort laundry, it would almost certainly struggle to clear a dinner table. Physical Intelligence, a startup focused on robotic foundation models, is attempting to breach this wall with its latest development, π0.7.
The model relies on a technical phenomenon known as compositional generalization. Rather than requiring a bespoke program for every conceivable action, π0.7 can synthesize disparate skills learned in other contexts to navigate unfamiliar scenarios. It is a process of creative assembly—taking the mechanical logic of one task and applying it to a problem the system has never technically "seen" before.
This approach mirrors the way the human brain functions when encountering a new environment. Humans do not require a manual for every new room they enter; they rely on a vast library of spatial and mechanical experiences to improvise. By imbuing robots with this same ability to generalize, Physical Intelligence moves the field away from specialized industrial tools and toward truly versatile agents.
While the technology remains in its early stages, the shift from narrow AI to generalized physical intelligence marks a critical inflection point. The objective is no longer just to build a robot that can perform a single task perfectly, but to develop a cognitive framework that can figure out how to do nearly anything.
With reporting from Exame Inovação.
Source · Exame Inovação
