Google is putting advanced artificial intelligence (AI) models into robots to give them an AI brain.
On Friday, July 28 Eastern time, Google DeepMind announced the launch of a new product for the field of Robotics – an AI model called Robotics Transformer 2(RT-2). It’s a new “visual-Speech-action” (VLA) model that can help train robots to understand tasks like trash throwing.
Based on the Transformer model, the RT-2 is trained on text and images on the Internet to directly instruct the robot to perform actions. Just as language models are used to train AI to learn ideas and concepts of human society through online text, RT-2 can also use online data to inform the robot of relevant knowledge and guide the robot’s behavior.
Google says, for example, that if we’re going to get old robotic systems to do the same thing, we have to explicitly train robots to know what trash is, and to pick it up and throw it away. RT-2 can pass relevant knowledge from the Internet to the robot, so that the robot knows what garbage is without explicit training, and even knows how to throw garbage even if it has never been trained to do so.
According to Google, RT-2 has the ability to translate information into action, and with it, robots are expected to adapt more quickly to new situations and environments.
Because after more than 6,000 robot experiments testing the RT-2 model, the Google team found that when faced with tasks that were already in the training data, or “seen” tasks, the RT-2 performed just as well as its predecessor, the RT-1. In the novel, never-before-seen mission scenario, the RT-2 nearly doubled its performance, achieving a success rate of 62 percent compared to 32 percent for the RT-1.
In other words, with RT-2, robots can learn as much as humans do, applying the concepts they learn to completely new situations.
According to Google, the RT-2 demonstrated the ability to generalize application and semantic and visual understanding beyond the robot data it was exposed to, including interpreting new commands and responding to user commands through basic reasoning, such as reasoning about categories and high-level descriptions of objects.
Google’s research has also shown that by combining reasoning with chains of thought, RT-2 can perform multi-stage semantic reasoning, such as determining which object can be used as a temporary hammer or which type of drink is best for a tired person.
Google has no immediate plans to release or sell RT-2 robots on a large scale, but eventually, the robots could be used in warehouses or as home assistants.
Vincent Vanhoucke, head of robotics at Google DeepMind, said: “RT-2 demonstrates not only how advances in artificial intelligence are quickly being incorporated into robotics, but also the great promise of more general-purpose robots.”
But Google’s first robot algorithm model with the ability to “self-learn” further heightens the concern that artificial intelligence is out of control, and whether humans have opened a Pandora’s box?
In the Terminator, robots’ ability to “self-learn” may allow them to gradually exceed the limits set by humans and achieve a level of intelligence beyond expectations. This transcendence could lead to robots with autonomous awareness and decision-making capabilities, which also means that they have the potential to pose unexpected threats to humans.
With the rapid development of artificial intelligence technology, technology giants and global regulators have become increasingly aware of the importance of artificial intelligence security issues, and have taken action to address this challenge.
Let’s hope AI goes the way of the Jetsons, where robot family members live in harmony with humans, rather than the Terminator.