Keep pulling the thread on Karol Hausman & Danny Driess.
The AI robotics company Physical Intelligence has raised over $400 million in funding.
The RT2 experiment, in which a robot was instructed to move a Coke can to a picture of Taylor Swift, was a pivotal moment that demonstrated the viability of large-scale AI models for robotics.
The RT2 model demonstrated the ability to connect abstract concepts learned from internet data, such as identifying Taylor Swift, with the physical motor commands required to complete a task, despite never having seen Taylor Swift in its robotics training data.
The success of the RT2 model showed that knowledge can be transferred from a pre-trained vision-language model to a robot with only a small amount of robotics-specific data, reducing the need to create an "internet of robot data."
Physical Intelligence's Pi Zero model, released within the company's first six months, demonstrated the capability to perform highly dexterous tasks such as folding laundry, building boxes, and bussing tables.
Physical Intelligence's Pi05 model demonstrated generalization by successfully performing long-horizon tasks, such as cleaning a bedroom or making a bed, in a novel home environment it had not previously encountered.
According to Carl Hausman, the common belief that robotics is bottlenecked by hardware is a misconception; the primary constraint on progress is the lack of sophisticated AI intelligence, not hardware capabilities.
A key finding in Physical Intelligence's Pi05 paper is that a robotics model trained on a sufficiently diverse set of environments can achieve the same level of performance in a completely new kitchen as a model specifically trained with data from that kitchen.
Danny Drace believes that solving general-purpose robotics is still fundamentally a research problem, not yet an execution or scaling problem.
Carl Hausman states that a key unsolved problem in robotics is establishing a scaling law that directly connects financial investment to model capability, which would turn the field into an execution problem.
Carl Hausman states that simulation is much harder for robot manipulation than for locomotion because the core difficulty is modeling the complex and unpredictable objects the robot interacts with, not just the robot's own limbs.
Carl Hausman believes the biggest risk for Physical Intelligence and the entire robotics field is not competition, but the fundamental scientific risk that the problem of physical intelligence proves unsolvable.