Physical Intelligence, a two-year-old robotics startup based in San Francisco, has unveiled new research showing its latest model can perform tasks it was never explicitly trained on. The model, called π0.7, demonstrates compositional generalization by combining skills from different contexts to solve unfamiliar problems.

The researchers showed the model an air fryer it had seen only a couple of times in training data and, with no coaching, it managed to cook a sweet potato reasonably well. With step-by-step instructions, it performed successfully, hinting at the potential for real-time improvement in new environments.

The findings suggest that robotic AI may be approaching an inflection point similar to what we saw with large language models, where capabilities begin compounding in ways that outpace the underlying data. However, while impressive, the model still struggles with complex multi-step tasks and requires significant prompt engineering for optimal performance.

“Sometimes the failure mode is not on the robot or on the model,” says researcher Lucy Shi. “It’s on us. Not being good at prompt engineering.” The team acknowledges that standardized benchmarks for robotics are lacking, making external validation difficult.

The researchers admit to being genuinely surprised by their own results, which could signal a significant step forward in AI but also raise questions about the knowledge sources and limitations of such models.