Despite the amazing advances in AI in recent years, robots remain stubbornly dumb and limited. Those in warehouses often follow precisely choreographed routines, without much ability to sense their surroundings or adapt on the fly. The few industrial specimens that can see and grasp objects only perform a certain number of actions with minimal dexterity due to a lack of general physical intelligence. A more capable robot could take on a much broader range of industrial duties, perhaps after a series of minimal demonstrations; in the case of π0, it will need enormous variability to move around and clean up messes in human homes.
The general enthusiasm for the progress of AI has already translated into optimism about major new leaps in robotics: Elon Musk’s Tesla is developing a humanoid robot called Optimus, which the businessman says will cost between $20,000 and $25,000 and will be able to perform most tasks. tasks in 2040.
The future of π0 is promising
Previously, the way to train robots for difficult tasks focused on training a single machine on specific tasks because the learning seemed non-transferable. Some recent academic work has shown that, with sufficient scaling and tuning, learning can be transferred between different assignments and robots. A 2023 Google project called Open X-Embodiment involved sharing robot learning between 22 different machines from 21 different research labs.
A key challenge pursued by Physical Intelligence is that the same scale of robot data is not available for training as for LLMs in text form. So the company has to generate its own data and devise techniques to improve learning from a more limited set. To develop π0, the company combined so-called vision language models, which are trained with both images and text, with diffusion modeling, a technique borrowed from AI image generation, to enable a more general type of learning. .
For robots to be able to perform any task that a person asks of them, this type of learning will have to be expanded considerably: “There is still a long way to go, but we have a number of examples that illustrate what is to come,” concludes Levine.
Article originally published in WIRED. Adapted by Alondra Flores.
#π0 #robot #care #household #chores
