Physical Intelligence's π0: a generalist robot model for different robots
In one sentence On October 31, 2024 startup Physical Intelligence presents π0, a vision-language-action model built on PaliGemma (3B) with a flow-matching action expert, pre-trained on about 10,000 hours of data from 7 robot configurations and 68 tasks, able to fold laundry and bus tables.
Physical Intelligence (pronounced "pi") is a startup (π0's authors include Karol Hausman, Sergey Levine and Chelsea Finn) with the goal of building a general software "brain" for different kinds of robots.
On October 31, 2024 it presents π0 (pi-zero), its first model. It demonstrates it by folding laundry straight out of the dryer, bussing a table, assembling boxes and bagging groceries: hard tasks because fabrics and objects are soft and unpredictable.
The twist: the same model is trained on data from different robots (single arms, dual arms, mobile manipulators) and can be used directly or fine-tuned for new tasks.
Companies
Physical Intelligence
Tools
π0, PaliGemma
Tags
Sources