In the fascinating of robotics, one quickly realises the intricacy hidden within seemingly straightforward tasks. Actions that we, as humans, perform without a second thought, encapsulate potentially infinite variables that are innately grasped by our cognitive machinery. Robots, however, are not naturally afforded this ability.
For this very reason, a significant portion of the robotics industry is dedicated to refining robots’ skills in repeatable tasks within structured environments. Fortunately, the recent years have witnessed a series of monumental advancements in robotic learning, setting the industry on an exciting trajectory towards more adaptable systems.
Shaping the Future: Google DeepMind’s Revolutionary Robotics Transformer (RT)
In the past year, Google DeepMind’s robotics team astounded the industry with the launch of their Robotics Transformer—RT-1. The system, which trained Everyday Robot systems to execute tasks like picking, placing, and opening drawers, demonstrated remarkable performance. Leveraging a database of 130,000 demonstrations, the system achieved a staggering 97% success rate across more than 700 tasks, according to the DeepMind team.
Taking the Leap: Introduction of RT-2
Now, the team is ready to unveil the next evolution of their innovation, RT-2. Vincent Vanhoucke, DeepMind’s Distinguished Scientist and Head of Robotics, elucidates how this system allows robots to transfer concepts learned from smaller datasets and apply them in various scenarios effectively.
Google explains, “RT-2 exhibits enhanced generalization capabilities and a deeper understanding of visual and semantic cues beyond the robotic data it was exposed to.” This comprehension extends to interpreting new commands, responding to user directives by performing basic reasoning, and even determining the most suitable tool for an unfamiliar task based on existing contextual information.
The Advantage of RT-2: A Glimpse into Real-world Applications
To provide a clearer picture of RT-2’s practical benefits, Vanhoucke shares an everyday scenario where a robot is tasked with disposing of trash. In traditional models, the robot needs to be explicitly taught to recognise what constitutes as trash and then trained to pick it up and discard it properly – a level of detail that is not highly scalable for systems anticipated to perform a wide variety of tasks.
However, RT-2 brings a revolutionary change to this dynamic.”Benefiting from its ability to harness knowledge from an extensive web data corpus, RT-2 is already equipped with a rudimentary understanding of what constitutes trash, and can identify it without the need for explicit training,” shares Vanhoucke. He adds that RT-2 can even comprehend how to dispose of the trash, despite never being trained for this specific action.
Vanhoucke concludes, “Consider the abstract nature of trash — what was once a bag of chips or a banana peel transforms into trash after consumption. RT-2, harnessing its vision-language training data, can intuitively grasp this concept and carry out the task.”
The Road Ahead: Enhanced Efficiency in Performing New Tasks
According to the DeepMind team, the efficiency rate of executing novel tasks has seen a significant leap from 32% to 62% in the transition from RT-1 to RT-2. This remarkable improvement points towards a promising future where robots, empowered by Google’s DeepMind’s innovative systems, can adeptly navigate an ever-expanding universe of tasks, both familiar and novel.
