Physical Intelligence, a robotics startup with a promising new robot brain, has made a groundbreaking discovery. Their latest model, π0.7, can perform tasks it wasn't explicitly trained on, showcasing an unexpected level of adaptability. This capability, known as compositional generalization, is a significant leap forward in robotics. It allows robots to learn from various contexts and apply that knowledge to new situations, potentially revolutionizing how we interact with machines.
What makes π0.7 truly remarkable is its ability to learn from minimal data. The model can understand and use an air fryer, a task it was never trained on, by combining fragments of relevant data with broader web-based pretraining. This suggests that robots might soon be able to adapt to new environments without extensive retraining, making them more versatile and cost-effective.
However, the researchers are quick to point out the model's limitations. It still requires step-by-step instructions to perform complex tasks, and its performance can vary depending on how the task is explained. This highlights the importance of prompt engineering in robotics, a skill that Physical Intelligence is continually refining.
The team's findings have been met with both excitement and skepticism. While they are surprised by the model's capabilities, some critics argue that robots still face significant challenges in generalizing tasks. The internet, a vast resource for language models, is not available to robots, and clever prompting alone cannot bridge this gap.
Despite these challenges, Physical Intelligence's progress is impressive. The company has raised over $1 billion and is valued at $5.6 billion, attracting significant institutional investment. The potential for a system based on π0.7 to be deployed in the real world is a topic of ongoing discussion, with the team remaining cautious about timelines.
The future of robotics is bright, and Physical Intelligence's work is a significant contribution to the field. As the company continues to innovate, we can expect to see more adaptable and versatile robots, bringing us closer to the dream of a general-purpose robot brain.
