Physical Intelligence develops robot brain capable of learning new tasks without prior training

Physical Intelligence has published research on its new robot model, π0.7, which can generalize skills to perform tasks it has not been specifically trained on.

The model demonstrates compositional generalization, allowing it to combine learned skills to solve new problems, contrasting with traditional rote memorization approaches in robotics.

In tests, π0.7 successfully used an air fryer after minimal training, showcasing its ability to learn from limited data and verbal instructions.

Researchers noted that the model's performance improved significantly with better task explanations, indicating the importance of effective communication in training.

While π0.7 shows promise, it is not yet capable of executing complex tasks from a single command and requires step-by-step guidance.

The team acknowledges the lack of standardized benchmarks for robotics, complicating external validation of their findings, but π0.7 matched the performance of previous specialist models in various tasks.

Physical Intelligence has raised over $1 billion and is valued at $5.6 billion, with discussions underway for further funding that could increase its valuation to $11 billion. The company remains cautious about commercialization timelines.