Technology has frequently been used to enhance and even supercharge human skills. We created the printing press to facilitate information exchange, the abacus (and later the calculator) to facilitate computation, and the aeroplane to facilitate travel between locations. We have created new methods of information processing in recent years, particularly in the field of machine learning, to power practical technologies like Search, Assistant, Maps, and more. Robot Transformer 1 (RT-1), which regrettably is neither an Autobot nor a Decepticon, was developed by researchers at Google Research and Everyday Robots as a technique to assist robots learn from one another and absorb massive amounts of data to increase performance.
According to Vincent Vanhoucke, head of robotics at Google Research, “earlier this year, we worked with Everyday Robots to demonstrate that integrating a powerful language model, such as PaLM, into a robot learning model could not only enable people to communicate with a robot, but also improve the robot’s overall performance.” “This language model allowed assistant robots to comprehend and carry out a variety of requests, such as ‘I’m hungry, bring me a food’ or ‘help me clean up this mess.
“Today, we’re employing the Transformer, a similar architectural framework to PaLM’s, to aid robots in generalising what they’ve already learned. It can learn, just like we do, from all of its aggregate experiences doing things like looking at and fetching snacks, rather than only understanding the words underlying a request like “I’m hungry, bring me a snack.”
For run-time inference to be effective enough for real-time control, the Robot Transformer 1 (RT-1) is developed to tokenize robotic input and output operations, such as camera feeds, task instructions, and motor commands. RT-1 demonstrated its ability to significantly improve generalisation across new tasks, objects, and environments after being trained on a 130,000-episode dataset of more than 700 tasks collected from an Everyday Robots fleet over a 17-month period. Its accuracy was also increased by observing other robots in action.
