Physical Intelligence says its π0.7 model helps robots do new tasks

In short: Physical Intelligence says a new AI model called π0.7 can help robots figure out some tasks they were never directly taught.

What happened

Physical Intelligence, a two-year-old robotics startup based in San Francisco, published research on a new model it calls π0.7. The company says the model can guide robots through tasks that were not specifically included in their training.

The key idea is what the researchers call “compositional generalization.” In plain terms, it means the robot can mix and match things it already learned, like using familiar building blocks to solve a new problem. The company says this is different from the usual approach, where a robot is trained like a student memorizing one exact set of steps for one exact job.

One example involved an air fryer. The researchers said their training data had only a couple of weak hints about air fryers, like a robot pushing one closed and another robot placing a bottle inside one. Even so, they said the robot made a reasonable attempt on its own and then succeeded after a person gave step-by-step spoken instructions, like training a new coworker on the spot.

The team also said results can depend heavily on how instructions are worded. One researcher said an early test had about a 5% success rate, which rose to about 95% after around 30 minutes of improving the phrasing.

Why it matters

If these results hold up, it could make robots more practical in real places, like kitchens, warehouses, or hospitals. Instead of retraining a robot from scratch for every new task, people might be able to explain what to do in plain language and get useful results sooner. The researchers also stressed that this is still research, not a finished product, and they said it is hard for outsiders to verify claims because robotics does not have widely used standard tests.

Source: TechCrunch AI