The DeepMind team, Google unit, presented significant improvements in the field of robotics aimed at increasing the speed, efficiency and safety of robots in real conditions. The main innovation is the Autort data collection system with the “robotic constitution” based on the “three laws of robotics” Isaek Azimov. This concept involves the integration of the visual model of the language (VLM) and the large language model (LLM) to adapt to unfamiliar conditions and determine the spectrum of suitable tasks.
autort is controlled using a “robotic constitution” containing a set of safe instructions for LLM aimed at avoiding tasks associated with people, animals, sharp objects and electrical appliances. To ensure additional safety, DeepMind equipped robots with the function of automatic stop when exceeding a certain threshold of force on their joints and introduced the physical emergency switch for manual disconnecting robots by operators.
In seven months, Google has launched a fleet of 53 Autort robots in four office buildings, spending more than 77,000 tests. Some robots were controlled by a remote man, while others worked according to a predetermined scenario or completely autonomously using the artificial intelligence model Robotic Transformer (RT-2).
The robots used in the tests have a utilitarian design and are equipped with a camera, a robotic hand and a mobile base. In each case, the system uses VLM to understand the environment and objects, and then LLM offers a list of creative tasks, for example, “put a snack on the countertop” and selects a suitable task for performing a robot.
DeepMind also presented Sara-RT, a new architecture of the neural network, designed to improve the accuracy and speed of the existing RoboCFORMER RT-2 model, as well as RT-Trajectory, which adds 2D contacts for the best robots of specific physical tasks, such as wiping the table .
Although to robots that can independently serve and perform household tasks, they are still far away, their development and training can be based in the future on a system similar to Autort.