Google's robotics team creates a ``Robot Constitution'' inspired by the ``Three Laws of Robotics''

The robotics team at Google's AI research division, Google DeepMind, has announced research results aimed at a future in which robots can perform tasks simply by issuing simple commands. The training data collection system ' AutoRT ' includes a 'robot constitution' based on Isaac Asimov's 'Three Laws of Robotics.'

Shaping the future of advanced robotics - Google DeepMind

Google's DeepMind 'Robot Constitution' will make sure its AI droids won't kill you - The Verge
https://www.theverge.com/2024/1/4/24025535/google-ai-robot-constitution-autort-deepmind-three-laws

' AutoRT ' is a system that collects more practical and diverse training data to achieve appropriate training to create robots that are useful in the real world. In a seven-month demonstration experiment, we controlled 20 robots simultaneously in various buildings and collected a diverse dataset of 77,000 trials across 6,650 unique tasks with a total of 52 robots. I did.

Robots come in a variety of shapes, but the robot used in this experiment was a simple one consisting of a camera, an arm, and a moving base. Based on the images captured by the camera, it created a visual language model ( VLM) understands the environment and visible objects. The large-scale language model (LLM) played the role of a decision maker, creating a list of creative tasks that the robot could perform and selecting the appropriate task to perform from there.

An example of a task completed is a robot wiping a table.

???? The system first uses a VLM to understand its surroundings. Next, an LLM proposes creative tasks for each robot and selects which one they should perform.

In evaluations over 7 months, AutoRT safely orchestrated as many as 20 robots simultaneously across various offices. pic.twitter.com/6IKNqEOIS5

— Google DeepMind (@GoogleDeepMind) January 4, 2024

Safety is important when operating robots in the real world. Isaac Asimov said, ``Robots must not harm humans.'' ``Robots must obey human orders unless they violate the above principles.'' ``Robots must protect themselves unless they violate the above principles.'' He demonstrated the ``Three Principles of Robotics'' in his own science fiction novel.

Inspired by these three principles, AutoRT provides decision makers, or LLMs, with a 'robot constitution' that robots must abide by when performing tasks. Furthermore, the safety regulations stipulate that ``robots must not attempt work involving humans, animals, sharp objects, or electrical appliances.''

However, even with these ``robot constitutions'' and safety rules, safety cannot be guaranteed, and AutoRT includes measures such as ``automatically stopping when the force applied to the joints exceeds a certain level.'' All active robots will be kept within sight of a human observer with a physical kill switch.

In addition, Google DeepMind announced `` SARA-RT, '' which further improves the efficiency of Transformer models, and `` RT-Trajectory, '' which supports generalization of robot movements by adding visual outlines to training videos. I am.