Sergey Levine, co-founder of Physical Intelligence, discusses their mission to develop robotic foundation models capable of controlling any physical robot for any task, aiming for general physical intelligence. The approach leverages insights from large language models (LLMs) to imbue robots with common sense and the ability to learn continuously, moving beyond specialized applications to unlock broad innovation in robotics.
Summarized by Podsumo
Physical Intelligence aims to build *robotic foundation models* that can control any physical system for any task, drawing parallels to the generality achieved by LLMs in language.
A core challenge is combining the *knowledge acquisition of generative AI* (like LLMs for common sense) with the *continuous improvement of deep reinforcement learning* to surpass human-level performance in physical tasks.
The concept of *Moravec's Paradox* is central, highlighting that tasks easy for humans (e.g., picking up a cup) are often hard for robots, but machine learning is shifting this by making data-rich physical tasks more tractable.
The 'Bitter Lesson' — letting machines learn from data rather than being explicitly programmed with human knowledge — is a key, though still controversial, principle for achieving true generality and improvement.
Success would lead to a 'Cambrian explosion' in robotics, radically *lowering the barrier for individuals and companies* to experiment with diverse robot form factors and applications, similar to the impact of personal computers.
"The fundamental goal of Physical Intelligence is to develop robotic foundation models that can control basically anybody's system to do any task."
"If you want to put together some cool new robotics applications, some cool new robotics idea, you can have a build a monstrous stack and you need to basically solve the intelligence problem. But if there is a solution that someone can build on top of, there's a foundation model that you can prompt that'll provide like basic functionality and then you can fine tune it a little bit or adjust it as some way to your application. Now, it actually makes it a lot more tractable for lots of people, lots of companies, lots of individuals to try all sorts of different things."
"The bitter lesson says that you should not program the machine to think the way you think it should think, but you should let it learn from data."