Google DeepMind presented the “Constitution of Robots”
DeepMind’s robotics team unveiled three new advances that will help robots make faster and better decisions. One includes a learning data collection system using a “Robot Constitution” that will ensure the safety of the robot’s actions for humans.
Google’s data collection system, AutoRT, can use the Visual Language Model (VLM) and Large Language Model (LLM) to understand the environment, adapt to unfamiliar settings, and make decisions about appropriate tasks. The Robot Constitution, inspired by Isaac Asimov’s Three Laws of Robotics, is described as a set of “safety-oriented prompts” instructing LLMs to avoid choosing tasks that involve people, animals, sharp objects and even electrical appliances.
For added safety, DeepMind programmed the robots to automatically stop if the pressure on the moving joints exceeds a certain threshold. The company also offered to use a physical emergency switch that human operators can use in emergency situations.
In seven months, Google deployed a fleet of 53 AutoRT robots in four different office buildings and conducted more than 77,000 tests. Some robots were controlled remotely by human operators, while others operated either scripted or fully autonomously using the Robotic Transformer (RT-2) artificial intelligence learning model. The robots were equipped only with a camera, a manipulator and a mobile base. “For each robot, the system uses VLM to understand the environment and the objects within its range of vision. The LLM then suggests a list of creative tasks that the robot can perform, such as “Put a snack on the table”, and plays the role of a decision maker, choosing the appropriate task,” the company said.
Other new DeepMind technology includes SARA-RT, a neural network architecture designed to make the existing RT-2 robot transformer more accurate and faster. The company also announced RT-Trajectory, which adds 2D contours to help robots better perform certain physical tasks, such as wiping the table.
DeepMind introduced the Robotics Transformer (RT-1) system in 2022, and RT-2 came out in 2023. RT-1 was used to train the Everyday Robot on more than 700 tasks. The system included a database of 130,000 demonstrations, which, according to the DeepMind team, led to successful completion of tasks 97% of the time. RT-2 allows robots to effectively transfer concepts learned on relatively small data sets to different scenarios.