Google released a new AI robot model equipped with the latest developments in large language models
Google released a new robot model Robotics Transformer 2 (RT-2), equipped with the latest progress in large language models , and can help train robots to understand tasks such as throwing garbage.
RT-2 is a “vision-language-action” model that can train robots to perform corresponding actions by feeding them information and images on the Internet. This makes robots smarter and gives them new abilities to understand and solve problems.
For example, if previous systems were expected to be able to throw away trash, they had to be explicitly trained to recognize trash, as well as pick it up and throw it away. RT-2 is able to transfer knowledge from large amounts of web data, it already knows what garbage is, and can recognize garbage without explicit training.
Google claims that the new model has almost doubled the performance of the robot compared to the first version, increasing the accuracy rate from 32% to 62%.
Vincent Vanhoucke, director of robotics at Google DeepMind, said: “Because of the explosion of generative AI, we have to rethink the entire research plan. A lot of things that were done before have completely failed.”
Ken Goldberg, a professor of robotics at the University of California, Berkeley, said that robots are still not as dexterous as humans and perform poorly on some basic tasks, but Google uses artificial intelligence language models to give robots new reasoning and improvisational skills , which is a promising breakthrough.
