Google DeepMind unveils advanced system for teaching robots new tasks
Google’s DeepMind team presented a new artificial intelligence algorithm RT-2, which allows robots to effectively transfer concepts obtained from relatively small data sets to different scenarios.
According to the developers, RT-2 demonstrates improved generalization abilities, as well as semantic and visual understanding beyond the data obtained by the robot. This includes interpreting new commands and responding to user commands by performing basic reasoning.
The system effectively demonstrates the ability to determine the best approach to solving a specific new task based on available contextual information.
As an example, the developers cited a scenario in which a robot is asked to throw away garbage. In many models, the user has to teach the machine to determine what is considered waste, and then teach it to pick up the item and throw it in the bin. This level of detail is not very suitable for systems that have to perform many different tasks.
Instead, RT-2 uses the Internet to gain knowledge. A large array of web data allows the algorithm to learn on the go to perform tasks that it was not able to do before.
According to the team, when moving from RT-1 to RT-2, the efficiency of performing new tasks increased from 32% to 62%.
