Google is training robots the way it trains AI chatbots

Google wants to improve its robots by implementing the Robotic Transformer (RT-2) AI learning model.

The company’s vision-language-action (VLA) model has been updated with RT-2. The model teaches robots to recognize visual and language patterns better to understand directions and determine what items will work best for the request.

Researchers tried RT-2 in a kitchen office with a robotic arm. They asked the arm to decide what would make a good makeshift hammer (a rock) and what drink to give a tired person (a Red Bull). They also told the robot to move a Coke can over to a picture of Taylor Swift. The fact that the robot is a Swiftie is good news for people.

The company said in a paper that the new model was trained on data from the web and robotics. It used research gains in large language models like Google’s own Bard and combined them with robotics data (like which joints to move). It can also follow directions written in languages besides English.

Researchers have been trying for years to give robots better reasoning skills to figure out how to live in the real world. James Vincent of The Verge said that real life is always messy. Robots need more instructions even if it’s easy for a person to do something. For example, cleaning up a drink that has been spilled. People know what to do by nature: pick up the glass, get something to clean up the mess, throw that away, and be more careful next time.

Before, it took a long time to teach a robot something. Researchers had to set each direction by hand. But with the power of VLA models like RT-2, robots can access a larger set of knowledge to figure out what to do next.

Last year, Google’s first step toward making better robots was to say it would use its LLM PaLM in robotics. It made a system called PaLM-SayCan, which combines LLM with physical robotics.

The new robot from Google is not perfect. The New York Times saw a live demonstration of the robot and said it got soda tastes wrong and thought fruit was the color white.

Depending on who you are, this news is either good or makes you think of the scary robot dogs from Black Mirror, inspired by Boston Dynamics robots. In either case, we can expect an even smarter robot next year. It might even be able to clean up a spill with little help.

The post Google is training robots the way it trains AI chatbots appeared first on AsumeTech.

from Technology - AsumeTech https://asumetech.com/google-is-training-robots-the-way-it-trains-ai-chatbots/