Physical Intelligence Introduces π0.7 — a Robot Brain Capable of Mastering Unfamiliar Tasks

Physical Intelligence, one of the most discussed players in robotics, has unveiled π0.7—an updated version of its model that serves as a "brain" for robots. The key achievement: the system now handles tasks it was never specifically trained on.

Physical Intelligence was founded in 2023 by a team of former researchers from Google DeepMind, Berkeley, and other leading laboratories. From the start, the company set an ambitious goal: to create a universal AI brain for physical robots—analogous to what GPT did for text processing. In its first year, the startup raised approximately 400 million dollars; investors include Khosla Ventures, Thrive Capital, and Amazon.

The first public model, π0, was released in late 2024. It could control manipulators and perform everyday tasks: folding laundry, washing dishes. But it was trained on specific skills—each action was demonstrated in advance and embedded in the model's weights.

π0.7 goes further. According to the company, the new model can transfer knowledge to tasks it has never been trained on.

In other words, it generalizes rather than simply reproduces learned patterns. This is what makes the step significant: not improvement within familiar scenarios, but the emergence of nascent "common sense" applied to the physical world. The company itself is cautious in its wording—calling this an "early but significant step" toward the long-sought goal: a universal robot that handles any physical task without separate training for each new scenario.

This is a fundamental difference from the current state of affairs, where an industrial manipulator can do exactly what it was programmed to do—and nothing more. Generalization has been one of the key challenges in robotics for decades. Teaching a robot to tighten a bolt on a conveyor belt is straightforward.

Teaching it to "figure out" a new task based on a general understanding of physics and space is quite another. It was precisely the successes of large language models that spurred the entire industry: it turned out that large-scale training on diverse data gives models the ability to generalize. Physical Intelligence applies the same principle to physical actions in the real world.

The architecture of π0.7 is not detailed in its specifics. The company has historically built its models using a diffusion approach, applying it not to pixels but to the space of the robot's actions: instead of selecting the next step from a fixed set, the model generates a movement trajectory, gradually refining it—much like how a neural network refines an image.

The question of how truly "unfamiliar" the new tasks really are remains open. The company has not yet published detailed benchmarks, and video demonstrations—a traditional marketing tool of robotics startups—do not always reflect real reliability in uncontrolled conditions. Nevertheless, the fact itself is important: a startup with serious funding and a world-class team is recording a measurable step toward generalization.

The race for a "brain for robots" runs parallel to the race for large language models—and the stakes are comparable. If in the coming years a truly universal model for controlling a physical body can be created, it will transform the entire automation industry: from logistics to surgery. π0.

7 is not yet the finish line. But it is one of the few real signals that the road to it exists.