Keep pulling the thread on Karol Hausman, Tobi Springenberg.
Physical Intelligence's PIO5 model, released in April, demonstrated the ability for a robot to operate in a new home it had never seen before.
Physical Intelligence has already deployed a robot in a commercial setting to perform a valuable task, a milestone they originally expected would take five years to achieve.
Physical Intelligence aims to build robotic foundation models capable of enabling any robot to perform any task.
Physical Intelligence uses Vision-Language-Action (VLA) models to leverage the common-sense world understanding from internet-pre-trained models, reducing the need for robots to learn everything from firsthand experience.
The PiStarO6 model from Physical Intelligence uses reinforcement learning (RL) from experience, where the robot collects data by executing a policy and receives reward signals from human supervisors.
Using reinforcement learning, Physical Intelligence's Pi Star 0.6 model increased task throughput by over 2x compared to the demonstration-trained base policy on tasks including box building, coffee making, and laundry folding.
A robot powered by Physical Intelligence's Pi Star 0.6 model was able to serve coffee for 13 hours continuously and fold laundry for four hours.
Physical Intelligence's value function can predict an impending failure, such as inserting a portafilter at the wrong angle, 30 to 40 steps before the failure occurs.
Physical Intelligence is intentionally avoiding focusing on a single vertical application to prevent becoming a narrow application company, as they believe the value of solving general physical intelligence is far greater.
Physical Intelligence's models have shown an unexpected ability to generalize across vastly different domains, including driving, coffee making, drone flight, and surgical robotics, using a single underlying model.
Physical Intelligence uses a transformer-based model architecture with up to several billion parameters for its robotics foundation models.
The PiStar06 model from Physical Intelligence demonstrates performance levels that are beginning to be suitable for real-world deployment.