Robotic tasks that appear simple to humans are often incredibly complex for robots due to the numerous variables involved. The robotics industry traditionally focuses on repeatable tasks in structured environments, but recent breakthroughs in robotic learning are paving the way for more adaptable systems.
Last year, Google's DeepMind team introduced the Robotics Transformer (RT-1), training its Everyday Robot systems to carry out tasks such as picking up items and opening drawers. This system utilized a database of 130,000 demonstrations, resulting in a 97% success rate for over 700 tasks.
Now, DeepMind unveils its next evolution - RT-2. As described by Vincent Vanhoucke, DeepMind's Distinguished Scientist and Head of Robotics, RT-2 enables robots to apply learned concepts from relatively small datasets to varying situations. This system demonstrates improved generalisation capabilities, semantic and visual understanding beyond the initial robotic data.
RT-2 displays advanced comprehension, capable of interpreting new commands, performing rudimentary reasoning about object categories, and executing tasks based on existing contextual information. For instance, when asked to discard trash, RT-2 identifies what qualifies as trash and how to dispose of it, without explicit training. This knowledge transfer ability makes RT-2 highly scalable for a diverse array of tasks.
Notably, RT-2's efficiency in executing new tasks has improved from 32% to 62% since its predecessor, marking a significant progression in robotic learning and adaptability.