Google DeepMind Creates Three Advanced Models To Develop Robots

The application of three models to robots. (Doc. Google DeepMind)

JAKARTA Google DeepMind, Alphabet's Artificial Intelligence (AI) company, is making new progress in the robotics world. They built three systems as the initial foundation of the robot.

The systems they create based on the Robotics Transformers (RT) model are AutoRT, SARA-RT, and RT Trajectory. These three systems are believed to be able to make robots work faster, either in making decisions or in navigating the environment.

AutoRT is a system that uses the potential of a large foundation model. This system is referred to as an important component for robots because it combines large fundamental models such as the Large Language Model (LLM) and robot control models, namely RT-1 and RT-2.

By using AutoRT, robot developers can improve robotics learning to train controls from multiple robots simultaneously. This system can direct robots to perform various tasks in various settings.

This system safely regulates 20 robots simultaneously, and a total of up to 52 unique robots, in various office buildings, collects diverse datasets consisting of 77,000 robotic trials in 6,650 unique tasks, said the Google DeepMind Robotics team. Meanwhile, Self-Adaptive Robust Attention for Robotics Transformers (SARA-RT) is an RT neural network architecture that is developed more sophisticatedly. This model is claimed to be 10.6 percent more accurate and 14 percent faster than the RT-2 model.

SARA-RT, which uses a new model refinement method, namely up-training, can make the robot model more efficient. In addition, robots that use this network architecture can make robots work faster according to their duties.

"We designed our system to be easy to use and hoped that many researchers and practitioners would implement it in robotics and others because SARA provided a universal recipe to accelerate Transformers," the team explained.

Another model that is no less important is RT-Trajectory, which is a model that adds visual lines to describe the robot's movements in training videos. The model will take a collection of training data and coat it with 2D trajectory sketches.

This trajectory in the form of RGB will provide practical visual instructions at a low level. In DeepMind's trial, RT-Trajectory managed to double the robot's performance with a success rate of 63 percent.

RT-Trajectory can create a trajectory by watching human demonstrations about the desired tasks, and even receiving hand-drawn sketches. And it can easily be adapted to different robotic platforms," concluded DeepMind Robotics team.