Physical Intelligence: Robot π0.7 Defies Expectations

Physical Intelligence and the π0.7 Model: An Unexpected Discovery

Physical Intelligence, a Bay Area startup, has recently published research that could mark a turning point in the development of artificial intelligence for robots. Founded two years ago, the company is already in the spotlight thanks to its AI model, π0.7. This model has demonstrated a surprising ability to perform tasks it was never explicitly programmed to accomplish.

Traditionally, the development of robotic AI relies on learning from data provided by developers. This data allows the AI to create logical connections and execute specific tasks. However, π0.7 has challenged this approach by practicing what researchers call "compositional generalization." This means it can combine skills acquired in different contexts to solve new problems, much like a human using prior knowledge to understand a new device.

The Air Fryer Experiment

One of the most striking demonstrations of this capability was conducted with an air fryer. In π0.7's training database, only two occurrences mentioned this device: one where a robot closed the fryer, and another where a robot placed a plastic bottle inside. Despite this, π0.7 successfully cooked a sweet potato in the fryer, guided by step-by-step verbal instructions.

Ashwin Balakrishna, a researcher at Physical Intelligence and a PhD student at Stanford, expressed his surprise at this achievement, stating that it was the first time he had been deeply surprised by the capabilities of a model he knew well. "My experience has always been that when I know the data well, I can pretty much predict what the model will be able to do. I'm rarely surprised. But these past few months have been the first time I’ve been deeply surprised," he said.

Promising but Cautious Results

Despite these encouraging results, Physical Intelligence remains cautious. In its publications, the startup refers to "early signs" of generalization and "initial demonstrations" of new capabilities. The π0.7 model is not yet capable of performing complex multi-step tasks from a single high-level command.

Sergey Levine, co-founder of Physical Intelligence and a professor at UC Berkeley, emphasized the importance of prompt engineering, which improved the success rate of the air fryer experiment from 5% to 95% after thirty minutes of adjustments. "You can't just tell it 'make me toast,'" he acknowledged, illustrating the current limitations of the model.

Currently valued at $5.6 billion, Physical Intelligence is in discussions for a new funding round that could nearly double this valuation to $11 billion. However, no commercialization timeline has yet been communicated to investors, highlighting the company's caution in the face of these promising but still research-phase discoveries.