Google Created 'Robot Constitution' to Ensure Future AI Droids Do Not Harm Humans

The DeepMind robotics team has announced three new developments that it claims will help robots make quicker, smarter, and safer judgments in the field.

One of these is a method for collecting training data with a "robot constitution" to ensure that your robot office helper can get you additional printer paper without mowing down a human coworker who happens to be in the path.

Google Developed 'Robot Constitution' To Ensure Future AI Droids Won’t Harm Mankind

AutoRT, Google's data collection system, can comprehend its environment, adapt to unexpected conditions, and decide on acceptable activities by combining a visual language model (VLM) with a large language model (LLM), as per The Verge.

The Robot Constitution, based on Isaac Asimov's "Three Laws of Robotics," is defined as a collection of "safety-focused prompts" advising the LLM to avoid jobs involving humans, animals, sharp items, or even electrical appliances.

DeepMind configured the robots to stop automatically if the force on their joints exceeds a particular level, and they have a physical kill switch that human operators may use to disable them. Google sent a fleet of 53 AutoRT robots into four separate office buildings over the course of seven months, doing over 77,000 trials.

Some robots were remotely directed by humans, while others worked from a script or entirely autonomously, utilizing Google's Robotic Transformer (RT-2) AI learning model. The experimental robots are more practical than spectacular, with merely a camera, robot arm, and mobile base.

"The system uses a VLM for each robot to understand its environment and the objects in view. Next, an LLM suggests a list of creative tasks that the robot could perform, such as 'Place the snack onto the countertop,' and acts as a decision-maker to choose an appropriate task for the robot to perform," Google explained in a blog post.

DeepMind Advances Robotic AI with SARA-RT Neural Network and RT-Trajectory for Physical Tasks

DeepMind's other new technology is SARA-RT, a neural network architecture meant to improve the accuracy and speed of the current robotic transformer RT-2. It also launched RT-Trajectory, which adds 2D outlines to assist robots in doing specific physical activities like wiping off a table.

We appear to be a long way from autonomous robots that serve beverages and fluff pillows, but when they arrive, they may have learned from a system like AutoRT.

Photo: Pawel Czerwinski/Unsplash