Google’s robots getting smarter with aid from AI language models

Google wants to make its robots smarter with the release of the AI learning model Robotic Transformer (RT-2). Google announced a new artificial intelligence model on Friday that can help it train robots to understand tasks like throwing out trash.

The Robotics Transformer 2 (RT-2) is a vision-language-action model trained on information and images from the internet that can be translated into actions for the robot, Google said in a blog post.

The model teaches robots to better recognize visual and language patterns to interpret instructions and infer what objects work best for the request. 

While a task like picking up the trash sounds simple to humans, it requires an understanding of a series of tasks for a robot to learn. For example, the robot must first be able to recognize what items constitute trash, then know to pick it up and throw it away.

Rather than program a robot to do those specific tasks, RT-2 allows the robot to use knowledge from around the web to help it understand how to complete the task, even if it hasn’t been explicitly trained on the exact steps.

The new model nearly doubled the robots’ performance on previously unseen scenarios, compared with the earlier version of the model, Google said. The new version can use rudimentary reasoning to respond to user commands, Google added.

The company doesn’t have imminent plans to widely release or sell robots with the new technology, The New York Times reported. But eventually, they could be used in warehouses or as home assistants, the Times added.