Google Drafts “Robot Constitution” to Ensure AI Robots Safe to Humans

Google’s DeepMind company recently announced three new developments, one of which is the drafting of a “robot constitution” for a system that collects training data to ensure that AI robots do not harm humans.
Google’s data collection system AutoRT can utilize both Visual Language Modeling (VLM) and Large Language Modeling (LLM) to understand its environment, adapt to unfamiliar environments, and determine appropriate tasks.
The “Robot Constitution” drafted this time is inspired by Isaac Asimov’s “Three Laws of Robotics”, focusing on safety and instructing LLM to avoid selecting tasks involving humans, animals, sharp objects, and even appliances.

DeepMind indicates that in order to improve safety, programming has restricted the joints of the robot to automatically stop when the force exceeds a certain threshold. In addition, it is equipped with a dedicated physical emergency stop switch.
Google stated that in the past 7 months, it has deployed a fleet of 53 AutoRT robots in four different office buildings and conducted over 77000 experiments.
Some robots are remotely controlled by human operators, while others run based on scripts or fully autonomously using Google’s Robot Transformer (RT-2) artificial intelligence learning model.
The robots used in the trial look more practical rather than flashy – only equipped with cameras, robotic arms, and a mobile base. For each robot, the system uses VLM to understand its environment and objects in its field of view. Next, LLM proposed a list of creative tasks that robots can perform, such as placing snacks on the table and playing the role of decision-makers, selecting appropriate tasks for robots to perform.
According to the report, another new technology of DeepMind is SARA-RT, which is a neural network architecture that can be more accurate and faster than the existing Robot Transformer RT-2.
Google also announced RT Trajectory, adding 2D contours to help robots better perform specific physical tasks such as wiping tables.