Sergey Levine, co-founder of Physical Intelligence, says the company aims to build robotic foundation models that can control basically any embodied system to do any task.
Levine argues general models may beat narrow ones in the long run, as language models did, and says real robot generalization means doing mundane tasks anywhere, not flashy demos in clean settings.
Levine says robots will be more like a toolkit than a single humanoid form, and notes robotic surgery is still done entirely through teleoperation.
Physical Intelligence bets one general brain can run any robot
Physical Intelligence is betting that robotics will follow language models, with one general model for many robot bodies beating narrow systems. The missing ingredient is a robot brain, and a model that understands physical interaction would make new applications far easier to build. Data is scarce and the timeline is uncertain, so practice and self-collected data matter.
Takeaways
General robot models should beat narrow ones. Physical Intelligence wants one foundation model that controls almost any robot for any task. Levine argues generality may prove easier long term, as with language models, which took over domain-specific applications because they draw on much broader data.
Real generalization means mundane tasks anywhere. Levine says impressive demos usually pick one cool task in a clean setting. True generalization is doing something ordinary in any environment. Public demos and social media videos often hide the true state of robotics.
The model starts from a language model. The approach is a vision-language-action model adapted from an LLM for robot control. It gains common sense through chain of thought and improves through practice, as the espresso robot did with reinforcement learning, gaining robustness, speed and throughput.
Hardware is simple and the model transfers. The robots use three cameras and no touch or force sensing. The same model worked on multi-fingered hands and robots with different degrees of freedom without changes. Levine says the transfer across embodiments surprised him.
Data is the open question, so robots must gather it. Robotics has no internet-sized dataset and nobody knows how much robot data is enough. Levine says systems should collect more data themselves, as Tesla does. Domains with easy data will get easy, while elder care and child care stay hard.
Robots will be a toolkit, not one humanoid. Levine compares the impact to personal computers, which unleashed many applications. Humanoids are one of many robot types, and the intelligence problem looks the same across bodies. Medicine is one opening, since robotic surgery is still fully teleoperated.
Learning versus programming is still contested. Levine says the bitter lesson is not universally accepted. The strongest counterargument is that reliable robots need textbook physics. Humanoid pipelines lean on simulation while manipulation work leans on real data, and he finds it surprising they look so different.
What could slow robots in homes
The breadth of situations a home robot meets is the biggest risk, and the long tail needs common sense.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Levine suspects a long tail of challenges in how technology and people interact will delay home robots.
Public acceptance shapes the timeline as much as technology does.
Useful deployment is a bootstrap challenge that must pass an activation energy threshold, so timing is uncertain.
Elder care and caring for small children will be harder for robots than people think.