Advancing Robotic Learning: Google’s DeepMind Unveils RT-2 System
Humans perceive robotic tasks as simple, but robots face inherent complexity due to their inability to consider various variables. However, robotic learning has significantly progressed, driving the industry towards adaptable systems. Last year, Google DeepMind’s Robotics Transformer (RT-1) garnered attention by achieving a 97% success rate in “over 700” tasks based on a database of 130,000 demonstrations.
Source: Google DeepMind
Now, Google introduced RT-2, a ground-breaking system empowering robots to transfer learned concepts to diverse scenarios with improved generalization capabilities and enhanced semantic and visual understanding. RT-2 empowers robots to apply knowledge from relatively small datasets across various situations effectively. The system’s exceptional capacity enables it to interpret new commands, perform rudimentary reasoning, and identify the most suitable tools for specific novel tasks based on contextual information.
A notable example involves a robot assigned to dispose of trash. Unlike previous models requiring explicit training, RT-2 draws insights from a vast corpus of web data, enabling it to recognize trash without dedicated instruction. Moreover, it infers the appropriate disposal method, making it a scalable and versatile solution for robots undertaking diverse tasks.
The evolution from RT-1 to RT-2 resulted in a surge in the efficacy rate for executing new tasks, reaching an impressive 62%. This significant improvement underscores RT-2’s potential to advance robotic capabilities, propelling the industry toward creating adaptable and efficient systems.
