Physical Intelligence has unveiled the AI model π0.7, which allows robots to accomplish previously unimagined tasks.

Physical Intelligence, a robotics startup from San Francisco, has launched a new AI model called π0.7, which allows robots to perform tasks for which they were not specifically trained. This represents a significant step towards more flexible robotic intelligence, capable of adapting to unknown situations. Unlike traditional methods of training robots, which focus on a single task, π0.7 combines learned abilities in different contexts to apply them to new challenges, a process known as compositional generalization.
An interesting example was the use of an air fryer, where the robot, without direct training data, managed to cook a sweet potato with the help of verbal instructions from a human. This suggests that future robots could improve performance in new environments through natural coaching, eliminating the need for costly retraining cycles.
Physical Intelligence compared π0.7 with previous robotic systems, demonstrating that the new model matched their performance in tasks such as making coffee and assembling boxes. Although researchers acknowledge the lack of universal evaluation standards in robotics, the results suggest significant progress.

The startup has already raised over 1 billion dollars and is discussing a new funding round, estimating that it will reach a valuation of nearly 11 billion dollars. This investor confidence reflects the belief that robotics could enter a phase of rapid progress, similar to that of generative artificial intelligence.