Secuzine

spot_img

Father of Molly Russell honoured with MBE for contributions to online child safety

Ian Russell acknowledged his contribution to online child safety through the establishment of the Molly Rose Foundation, reflecting on Molly's legacy. He emphasises the...

Aided by AI-Language Models, Google’s Robots Are Getting Smart

Imagine instructing a robot to “Pick up the extinct animal” and witnessing it effortlessly grab a plastic dinosaur. Until recently, this level of reasoning and adaptability was beyond the capabilities of robots. However, a quiet revolution in the field of robotics is now underway, driven by the integration of advanced language models powered by AI. Google, a pioneer in this field, has started equipping its robots with the equivalent of artificial brains, enabling them to perform logical leaps like associating “extinct animal” with a “plastic dinosaur.”

While robots have not yet reached human-level dexterity and still struggle with certain fundamental tasks, the use of AI language models by Google to enhance robots’ reasoning and improvisation skills is promising, according to Ken Goldberg, a robotics professor at the University of California, Berkeley. 

Google’s researchers came up with a novel idea: what if robots could learn new skills using an AI language model trained on extensive internet text, instead of being programmed task by task? Google’s initial attempt to merge language models and physical robots was the research project called PaLM-SayCan. However, it had limitations because the robots lacked the ability to interpret images, a crucial skill for navigating the real world.

Google’s latest robotics model, RT-2, addresses this challenge. It’s described as a “vision-language-action” model, meaning it not only perceives and analyzes its surroundings but also guides a robot’s movements. RT-2 achieves this by tokenizing the robot’s movements into numbers and incorporating these tokens into the same training data as the language model. Over time, similar to models like ChatGPT or Bard that predict the next words in a sentence, RT-2 learns to predict how a robot’s arm should move to perform tasks like picking up a ball or throwing an empty soda can into a recycling bin.

In a remarkable hour-long demonstration, RT-2 showcased its capabilities, successfully executing complex instructions like “move the Volkswagen to the German flag” by locating a model VW Bus and placing it on a miniature German flag.

This breakthrough opens up possibilities for using robots in various environments, from offices to homes, where physical tasks abound. As technology continues to advance, the synergy between AI language models and robotics promises to usher in a new era of smarter, more adaptable robots capable of tackling a wide range of tasks.