CITRIS researcher Sergey Levine’s startup, Physical Intelligence, is working to achieve the next breakthrough in AI: intelligent robots with human-like capabilities.
Having recently raised $400 million in investment for a total valuation of over $2 billion, the company aims to give robots human-like understanding of the physical world, and thus, nuanced functionality using vast amounts of sensor and motion data from robots performing a large variety of tasks. The startup is powered by the idea that a robot can be trained on three-dimensional information, much like ChatGPT, and other large language models (LLMs), are trained on written information. Levine and other cofounders have shown that LLMs, though unable to interact with the physical world, could provide the basis for this physical knowledge, containing extensive data on objects and actions.
Levine and his team have expanded their model to conduct more than 50 common household tasks, including unloading a dryer, cleaning a table, and folding clothing. Their next challenge involves feeding the model vastly larger amounts of data to master the complexity of navigating the physical world.
This research builds upon previous breakthroughs in robotic learning, including one in which Levine and his team used reinforcement learning to demonstrate a robot teaching itself to walk.