Founded by former employees of Google, Physical Intelligence launched model 0.7, which operates on the principle "vision-language-action". According to the creators, the system has made a significant leap in generalizing skills and successfully performs operations that were not specifically taught. During the tests, the robot with 0.7 mastered the unfamiliar control of the hardware design, carefully folded the t-shirts without having previously loaded scripts, and partially prepared the potato in an unknown aerogel.

The main innovation is that the machine can be given not only a goal, but also a way to achieve it. The net takes text descriptions, speed and quality parameters, as well as visual references in the form of pictures with expected results. This makes it possible to mix information from different sources and change tactics on the fly, without having to retrain. Such flexibility has long been regarded as the preserve of large language models and was virtually unattainable in robotics.

Although 0.7 is not always without detailed step-by-step instructions for solving difficult tasks, developers see it as a first step towards creating truly universal mechanisms that adapt to new environments without manual re-engineering.