Physical Intelligence: Robot π0.7 Defies Expectations

⚡

Key Takeaways

1Physical Intelligence revealed that its model π0.7 can perform unprogrammed tasks, such as cooking with an air fryer.

2The robot demonstrated a capacity for compositional generalization, combining skills to solve unknown problems.

3Despite an initial success rate of 5%, adjustments allowed for a 95% success rate with verbal instructions.

💡Why it matters — These findings could revolutionize robot autonomy, paving the way for unexpected and innovative applications.

An Unexpected Discovery at Physical Intelligence

In the ever-evolving world of artificial intelligence, surprises abound. Physical Intelligence, a Bay Area-based startup, has recently made a discovery that could be a game changer. Founded two years ago, this company is already in the spotlight thanks to its AI model, π0.7. This model has demonstrated capabilities that were not explicitly programmed by its developers.

The Compositional Generalization of π0.7

Traditionally, the development of AI for robots relies on a method where developers provide data that the AI uses to perform tasks. However, π0.7 has broken this mold by practicing what researchers call compositional generalization. This means that the model is capable of recombining skills acquired in different contexts to solve problems it has never encountered before. This ability is comparable to that of a human who, without ever having manipulated a specific device, understands how it works by analogy with what they already know.

A Demonstration with an Air Fryer

The most striking example of this capability was observed with an air fryer. In the model's training data, the fryer was mentioned in only two distinct contexts. In one, a different robot simply closed the device, and in the other, another robot placed a plastic bottle inside on human instruction. These occurrences were anecdotal and had no direct relation to the act of cooking. Yet, π0.7 attempted to cook a sweet potato in the device without any prior instructions. With step-by-step verbal instructions, it even succeeded in accomplishing this task.

Reactions from Researchers

Ashwin Balakrishna, a researcher at Physical Intelligence and a PhD student at Stanford, expressed his surprise at these results. He stated, "My experience has always been that when I know the data well, I can pretty much predict what the model will be capable of doing. I am rarely surprised. But these past few months have been the first time I have been deeply surprised."

Promising but Cautious Results

Despite these impressive results, Physical Intelligence remains cautious. In its research paper, the startup uses terms like "early signs" of generalization and "initial demonstrations" of new capabilities, emphasizing that these are research results and not products ready for deployment. Sergey Levine, co-founder of Physical Intelligence and a professor at UC Berkeley, explained that π0.7 cannot yet execute complex multi-step tasks from a single high-level command. "You can't tell it 'go make me toast,'" he acknowledged.

The Importance of Prompt Engineering

Prompt engineering, or the art of generating the right prompt to achieve the desired outcome, proved crucial in this experiment. An initial attempt with the air fryer had a success rate of only 5%. It took the teams thirty minutes to refine the wording of the instructions, resulting in a success rate of 95%.

A Growing Valuation

Physical Intelligence is currently valued at $5.6 billion and is reportedly in discussions for a new funding round that could nearly double this valuation to $11 billion. All of this, without having communicated any commercialization timeline to its investors. The next steps for this promising startup are eagerly anticipated.