Physical Intelligence's π0.7 model demonstrates unexpected robot generalization capabilities

On Thursday, Physical Intelligence, a robotics startup based in San Francisco and established two years ago, unveiled findings regarding its π0.7 model, which is capable of executing tasks without specific training. This model showcased compositional generalization, effectively addressing new challenges with minimal training data. Notably, it managed to operate an air fryer after only two training sessions. Co-founders Sergey Levine and Lucy Shi observed a performance enhancement from 5% to 95% with improved verbal instructions. The model achieved results comparable to specialists in activities such as coffee-making and laundry folding, though it still needs step-by-step directions for more complex tasks. The company has secured over $1 billion in funding and is currently valued at $5.6 billion, with potential discussions for a new funding round that could reach $11 billion.